L0 / SURFACE
Tanish Shah
Computer Engineering · Minor in Data Science · Mumbai
Engineer with a founder's bias.
Tanish's agent
Read transcript
Hi, this is Tanish's agent. He's not on the line, but I am. He builds the voice agents that make calls like this one, the automations behind them, and he's measured what happens on the GPU when a model answers. Stay on the line. Scroll down and I'll walk you through it.
Voice
I run the agents that make outbound calls for a global fintech-education company.
Calling coverage, 9 to 59 countries.
Regions shown are indicative.
Work · this layer
AI Intern, Marketing Automation & Growth
QuantInsti Quantitative Learning Pvt. Ltd., Mumbai
Jul 2026 to present
Make · LeadSquared CRM · OpenAI · Retell AI · Twilio · Vobiz · Claude Code
-
Outbound voice agents
- I own the Retell AI outbound calling agents on Twilio and Vobiz telephony.
- I added social-engineering guardrails to 28 live agent prompts.
- I fixed a silent failure in the 24-hour call-frequency guard.
- I built timezone-aware regional routing and a callback scheduler, and expanded calling coverage from 9 to 59 countries.
- I built a daily call-funnel report.
-
Voice Agent Call QA Auditor
github.com/RedPillCoder1/voice-agent-call-qa-auditor- A CrewAI multi-agent system that audits voice-agent calls, built on synthetic call data.
This is one turn of one call.
- 01 Telephony Twilio / Vobiz Carries the call audio in and out.
- 02 Speech to text transcribe Turns the caller's audio into text.
- 03 LLM decide what to say Reads the conversation so far and writes the next line.
- 04 Text to speech speak Turns that line back into a voice.
back to the caller
Orchestration
Behind every call and every email is a pipeline. I build them, ship them and pay for them.
Before QuantInsti
Real software for a real business.
Backend Automation Engineer Intern (sole technical hire)
AMAFHH Import and Export Pvt. Ltd., Mumbai
Jun 2025 to Sep 2025
FastAPI · Windows Server · Tally · Google Sheets
Google Sheets FastAPI XML Tally
- Owned the architecture, build and deployment of a production backend (HTTPS, IP allowlisting, HMAC-SHA256 request auth).
- Built a voucher pipeline from Google Sheets through FastAPI and XML into Tally, handling godown logic, conditional GST rules and a ±1% broker tolerance. It eliminated fulfilment errors on real transaction volume.
- Brought back in Oct 2026 to build a multi-user delivery-challan web app (per-user login, GST calculation, print preview, challan history).
Work · this layer
QuantInsti, continued
-
AI email engine, live on real leads
- A two-stage LLM pipeline (strategist, then writer) that writes one persona-targeted nurture email per CRM lead webhook, at about $0.33 per email.
- It ships through audit-driven releases.
- I designed its core rule: a lead's personal data steers the message but never appears in it.
-
Content engine
- AI short-form reels produced through a code pipeline.
- "The Desk" reached 2,300 views in under 48 hours against a 555-view account median. It became the account's best reel and now its reference format.
- Designed from a 44-reel baseline and a 25-account study.
- I also run an AI-assisted X replies programme and WhatsApp bulk campaigns.
-
Cost and reliability
- When the org hit its 300k-credit automation cap, I attributed spend per module and ranked the savings.
- I built a gate that stops malformed LLM output from reaching the CRM.
- I flagged security issues in live automation configs.
- I/O CRM lead webhook
A new lead arrives from the CRM.
- LLM Strategist
LLM. Reads the lead and decides what the email should say.
- LLM Writer
LLM. Writes the email. The lead's details steer it but never appear in it.
- GATE Output gate
Checks the output. Malformed output stops here.
- I/O CRM
Receives one finished email per lead.
one email per lead
A lead's personal data steers the message but never appears in it.
Inside the model
A model will say anything. My job is making sure only valid output gets acted on.
Work · this layer
-
Agentic Futures Trading Engine
github.com/RedPillCoder1/trading-bot- Turns natural-language trade intent into Binance Futures orders.
- A zero-trust validation layer blocks hallucinated parameters before any network call.
- Result: zero invalid exchange requests in paper-trading tests.
-
Autonomous Venture Lab
autonomousventurelab.vercel.app- An AI startup-investment analyst.
- A 4-agent LLM team (CEO, CFO, CMO, Critic) with schema-enforced JSON handoffs debates each venture.
- Then 10,000 Monte Carlo simulations per request across 4 exit scenarios return a probabilistic verdict in under a second.
-
4-agent software team
- PM, Developer, Tester and Reviewer agents that turn a feature spec into tested, reviewed code.
"short 0.01 BTC at market, 5x"
- t0 short
- t1 0
- t2 .
- t3 01
- t4 BTC
- t5 at
- t6 market
- t7 ,
- t8 5
- t9 x
| field | value | from | check |
|---|---|---|---|
| side | SELL | t0 | ok |
| symbol | BTCUSDT | t4 | ok |
| type | MARKET | t6 | ok |
| quantity | 0.01 | t1 t2 t3 | ok |
| leverage | 500 | t8 t9 | out of range · blocked |
blocked before any network call
The silicon
Every answer a model gives costs time on a GPU. I built a profiler that shows where it goes.
Flagship · open source
InferenceLens
Python · PyTorch · FastAPI · WebSocket · MPS
An open-source GPU-layer profiler for self-hosted LLMs. It covers the gap between LangSmith and NVIDIA Nsight, because neither shows per-layer GPU attribution.
It located 4x layer-level attention bottlenecks.
GPT-2 on Apple unified memory shows 33% p50 to p99 jitter. Llama 3.1 8B on a dedicated NVIDIA GPU, a model 64x larger, shows 7.8%.
GPT-2 · Apple unified memory
n = 153 · real requests · 25 Apr 2026
← 6 faster, below 0.8 s
- p50
- 1.0956 s
- p95
- 1.2367 s
- p99
- 1.4565 s
- p50 to p99
- 33% jitter
Latency ladder
Llama 3.1 8B · dedicated NVIDIA VRAM · n = 1,000 · summary statistics
- p50
- 2.3219 s
- p95
- 2.4489 s
- p99
- 2.5029 s
- avg
- 2.3296 s
- p50 to p99
- 7.8% jitter
Latency in seconds. Both axes cover the same range relative to their own median, so the bands compare directly.
request · 2026-04-25 03:41:02 · total 1.2132 s
Generation is ~92% of request time and ~11x the sentiment stage. Flagged on 167 of 176 requests.
Call ended · 0:17
Building something with AI? Let's talk.
Datasheet
- Education
- D. J. Sanghvi College of Engineering · B.E. Computer Engineering, Minor in Data Science · CGPA 9.11/10
- Recognition
- NPTEL Top 1% nationally, Python for Data Science
- AI & Agents
- LLM pipelines · prompt engineering · multi-agent systems · guardrails · OpenAI Agents SDK · CrewAI · RAG
- Voice & Telephony
- Retell AI · Twilio · Vobiz · ElevenLabs · HeyGen
- Automation & GTM
- Make · Zapier · LeadSquared CRM · webhooks · WhatsApp · outbound email · short-form video
- Engineering
- Python · JavaScript · SQL · FastAPI · Node.js · Next.js · PyTorch · Docker · Git · Claude Code daily