Money leaks through process gaps
Unfiled claims, missed reimbursements, unbilled work, late penalties. Nobody owns it because it is tedious.
An agent reads every document and drafts the claim. Your team approves.
TechDevs AI · Built by Prasad IyerThe work between your systems, handed to agents that show their work.
I build production AI agents for operations teams, on Claude, ChatGPT, Gemini or whichever model fits the job: voice, WhatsApp, email, creative, video and ads. Every figure traces back to its source, and anything that sends, spends or files waits for a human yes.
30 minutes with Prasad. You leave with a build plan, whether or not we work together.
agent ›
Sample run: /laytime-claims --batch q3-voyages. Read 214 documents: statements of facts, NORs, charter parties. Laytime computed for 37 voyages. 6 demurrage claims drafted, every figure cites page and line. Review gate: 6 claims waiting for ops approval.
Sample run: agent whatsapp-concierge --live. 128 conversations today, 11 handed to a person. Product cards pulled live from Shopify. Answers grounded in the brand knowledge base. Complaint detected, routed to support with the full thread.
Sample run: /ads-optimiser --accounts meta,google --window 7d. 42 ad sets scanned, 5 fatigued creatives flagged. Proposed: move 18% of budget from prospecting to retargeting. 12 new hooks drafted from last month’s winners. Review gate: budget change waits for your approval.
Sample run: brandops make "why RPA breaks on PDFs". Script written in the founder’s voice. Voice, avatar, b-roll and captions assembled. One approval screen before render. 1080×1920 reel rendered, captions burned in.
- AI skills written and in daily use
- 100+AI skills written and in daily use
- MCP servers built so businesses run operations from Claude or any AI agent
- 40+MCP servers built so businesses run operations from Claude or any AI agent
- agent types, one review gate model
- 8agent types, one review gate model
- agents can run on your infra, on your keys
- Your cloudagents can run on your infra, on your keys
Plugs into
ClaudeChatGPTGeminiClaude CodeAgent SDKMCPWhatsApp BusinessShopifySAPSalesforceZohoGmailOutlookMeta AdsGoogle AdsGA4
The problem
Six ways operations teams pay for work nobody owns.
If two of these sound like your week, the work is ready for agents. Not a chatbot bolted on top: agents inside the process, with your team approving what matters.
Headcount grows with volume
Every 2x in orders or tickets means hiring again, training again, and managing again.
Agents absorb the volume. People handle the exceptions.
RPA broke on anything unstructured
Bots handle clean fields. They fail on PDFs, emails and judgment calls.
The agent reads the messy inputs RPA could not, and flags what it is unsure of.
“We tried AI and it made things up”
Leadership will not trust output it cannot trace. Fair.
Every figure links to the page and line it came from.
The engineers are on the roadmap
Your in-house team could build it. Just not this year.
I ship it, then hand your team the code, the skills and the runbook.
Knowledge sits with one person
When the ops lead goes on leave, the process stops.
Their judgment becomes a written skill the agent follows every time.
What I build
Eight agents. One way of building them.
Each one is a set of written skills, a model, your systems as tools, and a review gate. Pick the one closest to your problem; most builds combine two or three.
Voice calling agent
Answers and places calls in a natural voice. Qualifies leads, confirms orders, books appointments and chases payments.
- Inbound: picks up every call, day or night, in your customer’s language
- Outbound: COD confirmations, reminders, renewals and collections
- Writes the outcome, recording and transcript into your CRM
Cloud telephonyHubSpotZoho CRMGoogle Calendar
Guardrail. Hands the call to a person the moment it hears frustration or a question outside its brief.
Text agent
A concierge on WhatsApp, SMS and web chat that knows your catalogue and your policies.
- Shows live product cards with price and stock from your store
- Tracks orders, handles returns and answers policy questions
- Detects a complaint and routes it to a person with the full thread
WhatsApp Business APIShopifyZendeskFreshdesk
Guardrail. Answers only from your knowledge base. Says “let me get someone” instead of guessing.
Email agent
Triages the shared inbox and drafts replies your team only has to approve.
- Sorts every message by intent, urgency and customer value
- Drafts replies with the order, invoice or policy it relied on attached
- Pulls POs, invoices and claims into your system of record
GmailOutlookSAPSalesforce
Guardrail. Nothing sends until a person approves, until you decide a category is safe to automate.
Creative designer agent
On-brand statics, carousels and banners from a one-line brief.
- Locks to your palette, type and logo rules, and checks every output against them
- Generates variants per audience and per placement size
- Exports ad-ready files with a brand-check report
FigmaCanvaGoogle DriveMeta Ads
Guardrail. Never invents people, products or claims. Real-photo slots go on a shot list.
AI video creation
One prompt to a finished reel: script, voiceover, avatar, b-roll and captions.
- Writes the script in your voice from an idea, a post or a product page
- Assembles voice clone, avatar and b-roll automatically
- Shows you one approval screen before anything renders
Voice cloningAvatar videoInstagram ReelsYouTube Shorts
Guardrail. Every word on screen is a word that is spoken. No invented statistics.
AI ads creator
Angles, hooks, scripts and variants for Meta and Google, built from what already wins.
- Mines reviews, comments and your best ads for angles that convert
- Writes hooks, primary text and scripts per angle and audience
- Produces static and video variants for every placement
Meta AdsGoogle AdsAd LibraryCustomer reviews
Guardrail. A category compliance check runs before a creative reaches your review.
AI performance marketer
Watches your Meta and Google campaigns around the clock and proposes the next move.
- Flags fatigued creatives, wasted search terms and budget leaks
- Proposes budget shifts and bid changes with the numbers behind them
- Writes the weekly report your leadership actually reads
Meta Marketing APIGoogle Ads APIGA4Looker Studio
Guardrail. Spend changes wait for your approval. Caps and one-click rollback are built in.
Document and claims agent
Turns PDFs, scans and email attachments into structured data and drafted claims.
- Reads invoices, contracts, statements of facts and bills of lading
- Computes what is owed and drafts the claim or the ledger entry
- Cites the page and line behind every number
SAPOracleSharePointExcel
Guardrail. Low-confidence fields are flagged for a person, never guessed.
Work
Built, running, and honestly labelled.
Client names are withheld. Every card says what stage it is at: live, building, or scoped. Two of these run our own business, because we use the method before we sell it.
D2C gourmet food brandLive
An AI concierge that sells from the catalogue
The brand wanted shoppers answered on WhatsApp and the website at any hour, without a support hire.
- Concierge agent on a curated knowledge base
- Live Shopify product cards inside the chat
- Complaint detection that routes to the team
- Deployed on the client’s own AWS account
WhatsApp Business APIShopify Storefront APIVector knowledge baseAWS
What it provesA commerce agent that answers, recommends and sells around the clock, running on the client’s own infrastructure.
Multi-channel ordering platformActive build
Agents that write orders back into a live store
Orders arrived across channels and were re-keyed into the store by hand.
- Multi-channel order intake and routing
- Zone exclusivity and service filtering rules
- Order write-back into Shopify
- Notification flow and an operations dashboard
Shopify Admin APIOrder rules engineDashboard
What it provesAgents writing into a live commerce system, not just reading from it.
Our own productLive, our product
Agentic Outreach: signal-driven LinkedIn and email
Founder-led B2B teams need warm conversations, not spray-and-pray sequences.
- Intent-signal agents: post engagers, funding, hiring and news
- Warm-lead queue with copilot and autopilot-with-approval modes
- Its own MCP server, so campaigns run from Claude, ChatGPT or any AI agent
- Grounding checks that stop the agent inventing facts about a prospect
ClaudeMCPLinkedInEmail
What it provesWe run agents on ourselves first. The grounding checks exist because we caught the failure.
Our own agencyLive, used daily
TechDevs Agency OS
Running SEO and delivery for many clients without the work living in one person’s head.
- Audits, blog writing, backlinks and reporting run as agent jobs
- A reviewer on every job before anything reaches a client
- A client portal where sign-off is recorded as its own fact
- Cost tracked per client and per run
Claude Agent SDKJob queuePostgresReview gates
What it provesThe same method we sell, used every day on our own revenue.
Founder-led brandShipped, in use
One prompt to a finished reel
Publish video every week without a shoot, an editor or a studio.
- Idea to script in the founder’s own voice
- Voice clone, avatar and b-roll assembled automatically
- One approval page before render
- Captions that only show words actually spoken
ClaudeVoice cloneAvatarAuto-edit
What it provesAn AI video pipeline with a human gate, built and used by the person selling it.
Global agribusiness, shared-services centreScoped
A demurrage and laytime claims factory
Claims go unfiled because computing laytime from vessel paperwork is slow and tedious.
- Document reading for statements of facts, NORs and charter parties
- Laytime and demurrage computed per voyage
- Drafted claims with page-and-line citations
- An exception queue for the operations team
Document AIClaims engineReview queue
What it provesEnterprise-grade claims automation, designed for a GCC team that answers to audit.
Performance marketing agencyIn design
A performance marketing OS with an AI creative studio
An agency running a portfolio of client ad accounts wanted creative and reporting to scale without hiring.
- Brief to angles, scripts and AI shots
- Human gates at angle, storyboard and final cut
- Campaign analytics with an ad report card
- A decision log and a client portal
Meta AdsGoogle AdsAI videoApprovals
What it provesCreative, media buying and reporting designed as one loop with approvals.
Receipts, not vibes
“We tried AI and it made things up.” This is built to answer that.
Click any highlighted figure. Every number an agent writes carries a pointer to the document, page and line it came from, so a reviewer checks it in seconds instead of redoing the work.
Sample output · demurrage claim, voyage 31
The vessel tendered notice of readiness at 06:10 on 14 August1. Laytime commenced six hours after NOR2. Discharge completed at 18:40 on 22 August3. Allowed laytime was 4 days 12 hours4. Demurrage due: 2 days 3 h 50 m at $18,000 per day, $38,8755.
Every number cites its sourceDocument, page and line, one click away.
A human approves what sends, spends or filesReview queues are part of the build, not an afterthought.
Runs in your cloud, on your keysCompute and tokens billed to you directly. Your data stays yours.
Every run is loggedInputs, tool calls and outputs, replayable when someone asks why.
Prasad IyerCo-founder · CTO, AI & operationsProfile
The person behind the agents
Prasad writes the playbooks the agents run.
Prasad Iyer co-founded TechDevs and leads its AI practice. He works with AI agents every day, has built more than 40 MCP servers for businesses and has written more than a hundred skills: tested, reusable playbooks that turn a judgment call into something an agent does the same way every time.
Before TechDevs took all his time, he spent two and a half years as an applications developer at Oracle, on machine learning projects top banks used for model governance. That is where the habits come from: every figure traceable, every release approved. The SEO audits, site builds, proposals, outreach and video TechDevs delivers come out of systems he built and still runs.
Skills100+ written playbooks: SEO audits, site builds, proposals, video, outreach, code review. They run in Claude, ChatGPT, Gemini and other agents.
Parallel subagentsSpecialist agents fan out across isolated worktrees; a reviewer checks the merge.
40+ MCP serversYour systems exposed as tools any AI agent can call, from Claude to ChatGPT to Gemini, scoped, with no send button unless you add one.
Agent SDK workersLong-running jobs on a queue, each with its own reviewer.
Hooks and hard rulesWhat an agent may never do, enforced in code before it tries.
Review loopsEvery change loops through automated review until it comes back clean.
automate-seoseo-audit-reportbest-webdevbusiness-site-builderproposal-deckonboard-clientreviewloopblog-image-gengmb-optimizercrawl4ainano-bananasmooth-landing-pageseo-rank-analystseo-link-intelseo-backlink-prospect-finderseo-content-scorerseo-duplicate-analystseo-report-narrator
seo-page-rebuilderseo-outreach-prospectorseo-ai-visibilityseo-blog-writerseo-crawl-workerwebdev-strategistwebdev-copywriterwebdev-section-builderwebdev-svg-artistwebdev-video-producerwebdev-image-directorwebdev-site-analystwebdev-qanew-featurecode-structureevidence-driven-testingbefore-and-afterseo-technical
A slice of the 100+ skills and subagents in the working library.
How a build runs
From one person’s judgment to an agent’s routine.
- Step 01
Sit with the team
I shadow the people doing the work and write down every judgment call they make.
- Step 02
Write the skills
Each judgment becomes a skill with examples and edge cases. Your ops lead signs it off.
- Step 03
Build with gates
Agents run on the skills. Anything that sends, spends or files waits in a review queue.
- Step 04
Pilot on real volume
We run last month’s real inputs side by side with your team and measure accuracy before go-live.
- Step 05
Hand over the keys
Code, prompts, skills and runbooks are yours. A retainer keeps it running and improving.
Business ideas
Automations worth building this quarter.
Pick your industry. Each idea is a real pattern I would scope on a first call, with the agents it needs.
COD confirmation caller
Calls every cash-on-delivery order before dispatch and holds the ones the customer will not accept.
WhatsApp concierge with product cards
Answers sizing, stock and shipping questions and sells from the live catalogue.
Creative refresh engine
Spots fatigue in your ads and ships new on-brand variants before ROAS slides.
A lead caller that rings in under a minute
Calls every portal lead, qualifies budget and location, and books the site visit.
Site-visit follow-up on WhatsApp
Sends the brochure, floor plans and payment plan, then answers the questions.
Listing video generator
Turns photos and a floor plan into a walkthrough reel for every unit.
Appointment and reminder agent
Books, reminds and reschedules on calls and WhatsApp so the front desk stops chasing.
Insurance pre-auth pack builder
Pulls the right documents from the record and assembles the pre-authorisation pack.
Patient FAQ assistant
Answers only from clinician-approved content and escalates anything clinical.
Demurrage and detention claims factory
Reads the voyage paperwork, computes what is owed and drafts the claim with citations.
Shipment status agent
Answers “where is my shipment” by email and WhatsApp from your tracking feeds.
Invoice to PO matching
Matches supplier invoices to POs and receipts and queues only the mismatches.
Signal-based outbound
Finds prospects showing intent and drafts outreach that cites only verified facts.
Proposal generator
Turns discovery-call notes into a priced, on-brand proposal for review.
Client reporting autopilot
Pulls the numbers, writes the narrative and drafts the monthly report.
Accounts payable exception desk
Clears the routine invoices and hands people only the real exceptions, with reasons.
Vendor master data clean-up
Finds duplicates, missing tax IDs and stale bank details, and proposes fixes.
Policy answers for the ops floor
Turns the SOP binder into an assistant that cites the clause it answers from.
RFQ reader and quote drafter
Reads incoming RFQs, checks specs against your catalogue and drafts the quote.
Dealer orders on WhatsApp
Dealers order in a chat; the agent writes the order into your ERP.
Service call triage
Takes the breakdown call, captures the fault and books the right engineer.
Admissions counsellor agent
Calls and messages every enquiry, answers fee and course questions, books the counselling slot.
Course content to reels
Turns lessons into short videos for Instagram and YouTube every week.
Fee reminders and collections
Polite, persistent reminders by call and WhatsApp, logged against each student.
Fit check
Is this a fit? Tick what is true.
I take a small number of builds at a time, so I would rather tell you now if this is early for you.
0 of 5 true
Probably early.
The skills and ebooks in the resources library are a better place to start.
Browse resourcesWays to work together
Start small. Keep what works.
Scoping call
30 minutes on one workflow. You leave with a build plan and a fixed price, in writing.
Book it →Pilot build
One workflow, real volume, measured accuracy. Fixed scope, fixed price, review gates included.
Talk scope →Run and improve
Hosting, monitoring, new skills and fixes on a monthly retainer, three-month minimum.
Ask about retainers →Do it yourself
Skills, agents and ebooks from my own library, ready to install in Claude, ChatGPT, Gemini or your own agents.
Browse resources →Resources · AI skills and agents
Rather run it yourself? Take the playbooks.
A basic prompt gets you a draft. These skills get you the deliverable: the steps, scripts and quality checks Prasad wrote while doing the same work for paying clients, packaged so Claude, ChatGPT, Gemini or any agent you use follows them every time.
Skills, agents and ebooks.
Install the same playbooks I use on client work, singly or as bundles.
FAQ
What buyers ask before the call.
Will the agent make things up?
It is built not to. Agents answer from your documents and data, every figure carries a citation back to its source, and fields the model is unsure of are flagged for a person instead of guessed. On our own outbound agent, grounding checks block any message that states a fact about a prospect the research did not find.
Where does it run, and who owns it?
In your own cloud account if you want it there, with compute and model costs billed to you directly. You own the code, prompts, skills and runbooks.
Do you only build on Claude?
No. I pick the model per step: Claude for long documents, careful reasoning and tool use, ChatGPT or Gemini where they do a step better or cheaper, and small models for narrow, high-volume work. Skills and MCP servers are model-agnostic, so you are never locked to one vendor.
What does it cost?
Pilots are fixed scope and fixed price, followed by a monthly retainer to run and improve them. The number depends on volume and the systems involved, and you get it in writing after the first call.
What if Prasad is busy?
I design and review every build. A team I hired and trained maintains and extends it, and every agent ships with a runbook so the process never depends on one person, including me.
Can you work with our in-house engineers?
Yes. Many teams want the first agent built fast and then owned internally. I pair with your engineers, ship in your repository and leave the skills documented.
We are smaller than your usual client. Is there anything for us?
Yes. The skills, agents and ebooks in the resources library are the same playbooks I use on client work, packaged so you can install them yourself.
