Founding Engineer · LLM Evals & Guardrails · Regulated Fintech
Dhruv Puri
Founding engineer at GoReach Finance. Shipped an AI financial assistant to 2,000+ users, with the evals and guardrails that keep it safe with money.
Millions of impressions a month from pipeline-generated ad creative. 0.02 mm positional precision. 10M+ transactions validated. 20 projects shipped.
Operator file
REF A-01 · Updated 2026Founding engineer at GoReach Finance, where I built Reacher, the AI financial assistant on a SEBI-regulated advisory platform with $10M (₹85 Cr+) in AUM. Reacher has 2,000+ users, and I also built its evals and guardrails and the tools the platform's advisors use. Before this I was the founding AI/ML engineer at JuiceLabs, where I built an end-to-end video generation pipeline — idea in, finished cinematic ad out — shipping at millions of impressions a month for enterprise clients. Earlier, researchers used software and edge vision I built to 3D-print tissue for organ regeneration studies. Based in Bengaluru, India, and open to full-time and contract roles, remote or with relocation anywhere.
Exceptional problem-solver, serving skills with remarkable innovation. Never misses an opportunity.
Service record
5 roles · 3 cities · Chronological descFounding Engineer
- Built Reacher, the AI financial assistant on a SEBI-regulated advisory platform with $10M (₹85 Cr+) in AUM, and shipped it to 2,000+ users on iOS and Android. It answers money questions, finds and compares funds, explains a user's own portfolio and reads the statements they upload. When a question needs personal advice, it hands the conversation to a registered human expert.
- Built the evals Reacher runs on every change: checks that it stays on the right side of SEBI's line between education and advice, output-safety checks, and a scenario suite that replays real conversations against the live service and grades each reply against what the system did.
- That replay suite caught the hardest bug: Reacher sometimes told a user it had connected them to an expert when no handoff had happened. The reply read well, so every text-quality check passed it. I built a guard that compares each action a reply claims with the actions recorded for that turn and blocks the ones that never happened. It ran in shadow mode first so I could see how often it fired, and now runs on every reply.
- Built Reacher's guardrails in code: a crisis path in English and Hinglish, output-safety checks, and deletion of a user's data on request under India's DPDP Act.
- Built the multi-provider model routing behind Reacher, which falls back to another provider when one fails and cut its model costs 64%.
- Designed and launched our second AI product, an agent advisors use across their whole client book. It explains a client's portfolio, briefs the advisor before a conversation, answers questions across the book and drafts replies. Every figure comes from the platform's own data, and when a read comes back incomplete it says so. Nothing reaches a client without the advisor: drafts land in their usual composer and they decide what to send.
- Built the AI behind advisors' client calls: live and recorded-call transcription that handles Hinglish, chosen after testing several speech vendors; meeting summaries where code checks every quote against the transcript before an item is kept; and an LLM-as-judge call-quality review that scores calls against the firm's policies and cites the lines it relied on.
- Built the platform's notification system from scratch: event-triggered and scheduled messages plus admin-run campaigns across push, email, SMS and WhatsApp.
- On the customer React Native app I built the wealth and portfolio screens, return charts against market indices, the in-app AI chat and push notifications.
- Built our internal ops agent, which works across Slack, ClickUp and Google Workspace. It tracks tasks, schedules calls, takes meeting notes and posts a daily fintech news brief. Anything that sends or changes something waits for a person's approval, and it switches model providers when one goes down.
- Created Gist by GoReach, a daily, fully automated market newsletter with 2,000+ subscribers. It pulls financial news from many sources, fact-checks each story and matches every figure against the original article before it goes out. An issue that fails a check is held for a person to review.
- Wrote the team's engineering standards and the shared Claude Code setup everyone installs, led reviews of large changes from teammates, and planned the platform's move to Kubernetes on AWS.
Founding AI/ML Engineer
- Built the end-to-end video generation pipeline: idea → script → storyboard → finished cinematic video, length-agnostic.
- Multi-phase agentic architecture: granular per-atom user review, parallel multi-model orchestration, cross-family LLM-as-judge evaluation.
- Shipped 15-second ad creative for enterprise clients at millions of impressions a month.
- Ran weekly Claude Code enablement sessions for a 20-person team; advised enterprise clients on AI adoption.
Founding Software Engineer
- Built the human-machine interface for biomedical 3D printers; researchers used it to print tissue constructs for organ regeneration studies.
- Trained a YOLO model on 1,000 nozzle images, running locally on a Raspberry Pi, to drive auto-calibration at 0.02 mm positional precision; real-time OpenCV print QC.
- Cut user setup time 30%; improved print accuracy 50%.
SDET
- Automated device testing in Python, halving test cycles; validated Kafka pipelines and REST APIs across 10M+ transactions.
- Surfaced 50+ critical bugs pre-release, cutting post-release issues 20%.
Product Analyst
- Revamped the React frontend: +20% traffic, −40% production bugs; Cypress automation lifted coverage 35%.
Project register
P-001 — P-020 · Selected 9 + shelf
Video Generation Pipeline
Idea in, finished cinematic video out. Multi-phase agentic pipeline with per-atom review gates and cross-family LLM-as-judge. Millions of impressions a month.
Syto Studio
Human-machine interface for biomedical 3D printers, used in organ-regeneration research. YOLO on a Raspberry Pi for 0.02 mm auto-calibration, plus computer-vision QC.
Agentic RAG for Natural-Language SQL
Conversational agent turning natural language into safe SQL: memory, error recovery, injection protection.
Policy Compare
AI-powered policy comparison that helps regular people understand 50-page loan documents. Built after watching my parents lose hours to bank PDFs.
AgentVC
AI-powered pitch practice platform where founders rehearse against a simulated investor. Live demo shipped, insights from 50+ real investor conversations baked into the feedback logic.
Parameter Golf Autoresearch
OpenAI Model Craft Challenge: built an autonomous research agent that runs hyperparameter sweeps, scores candidates against the frontier, and decides what to test next. RTX 4060 for iteration, 8xH100 via RunPod for finals.
Advisor AI Agent
GoReach's second AI product: an agent financial advisors use across their whole client book. Launched November 2026.
Reacher
The AI financial assistant on GoReach's SEBI-regulated advisory platform, used by 2,000+ people. Evals that grade replies against what the system did, guardrails in code, and routing that cut model costs 64%.
Agent Memory
A chat-first finance agent, built on a hand-rolled Anthropic SDK loop with no framework, that carries what it learned in one conversation into the next one three days later. It ships with a behavioural eval.
Toolbox
Grouped by domain · Depth inferred from the work
Agentic Systems
The core practice: orchestrating models that plan, act, and get judged.
Regulated Systems
Regulated money. Every write leaves a trail.
Multimodal & Speech
Pixels, frames, and voice as one pipeline.
Infra
The substrate everything above runs on.
Consumer Apps
Where the work meets a user.
Say hello
Interested in new opportunities, short scope advisory, or a genuine conversation about AI engineering. Send a short note with what you’re working on, and I’ll reply within a day.