I build AI systems that survive contact with the real world.

Scaler AI Labs

AI Engineer, Internship · clients: OpenAI, xAI, Meta

Built reinforcement-learning environments and LLM infrastructure delivered straight to frontier-lab clients.

• Delivered 5 LLM agent environments (Shopify Admin, HelloSign, Asana MCP, Hootsuite, MS Excel). I drove the requirement calls myself to turn vague briefs into scoped specs, then hit hard external deadlines. Asana MCP validated at a perfect 1.0 across 52 tasks; HelloSign cleared client QA at 92.94%.
• Built Pulse, a production LLM extraction pipeline (desktop agent → S3 → Celery/Redis → LLM → ClickHouse) live for a 20,000-user org. On client request I re-cut daily roll-ups to hourly and shipped per-org system prompts, so their team could tune extraction without a code change.
• Cut LLM spend on that pipeline with a Hamming-distance filter that dropped near-idle frames before the model, and gated quality in CI, where Claude and Groq score every new environment on its PR.
• Built a distributed MCP server from scratch for multi-step tool-use workflows, plus a crawler pipeline (Playwright, Redis) generating training data at 100K+ pages/day.

Wealthy

SDE-1, promoted from SDE Intern in Mar 2025 · Fintech

Owned the internal CRM used daily by 7,700+ partners, sitting behind ₹4,000+ Cr in assets under management.

• Sat with the PM and ops users, turned their workflow complaints into shipped changes, and lifted conversion 37%.
• Tuned GraphQL resolvers and backend integrations to cut API latency 32% under peak load.
• Built the CRM's first version from an empty repository as an intern (React, Next.js, REST); the component and API patterns I set were adopted by the full-time team.

HuggingFace TRL

Three merged pull requests in the library most people use to train models with reinforcement learning.

• #6534 fixed a silent training-corruption bug: GRPO accepted a reward function returning the wrong number of rewards, so a single value was broadcast across every completion and quietly poisoned training. I filed issue #6533 with a reproduction and root cause, shipped the guard plus a regression test, then replicated the fix in RLOO on maintainer review.
• #6503 turned on the importance-sampling-corrected KL gradient by default. A maintainer pushed back, so I reproduced their script, found my own "this is a no-op" claim was wrong (it moves the gradient about 21% on-policy), said so, and the change was approved.
• #6477 completed the GRPOConfig.loss_type documentation across all eight loss types.

Python · TRL · GRPO All PRs

entangle

My own tool: a Go CLI that hands a live coding-agent session to another machine, or another person, with every file path rewritten so it actually resumes.

• Run send, read out three words, and the other side runs receive. The transfer is end-to-end encrypted and direct, with no account and no server in the middle, and secrets are scrubbed before anything leaves the machine.
• Sessions are pinned to absolute project paths, so the hard part is recovering each project's true path and rewriting every stored path for the target OS. Works with Claude Code, Codex, and opencode.
• 16 releases, a Homebrew tap, a VS Code extension, one-line install, and self-update. Fixed a user-reported Gatekeeper block and a zip-slip vulnerability, both covered by regression tests.

Go · GoReleaser · GitHub Actions Writeup Site GitHub

EleutherAI & Unsloth

Two open pull requests in the evaluation and fine-tuning stack.

• lm-evaluation-harness #3950: the 67-subject TMMLU+ benchmark stopped loading entirely on datasets >= 4 because it pointed at a script dataset. Repointing it at a parquet-native mirror fixes 66 subjects in one line; I proved parity by diffing git blob hashes across all 267 files.
• unslothai/notebooks #306: added a SmolLM3 conversational fine-tuning notebook.

Python · LLM evaluation · fine-tuning

Viveka

A reinforcement-learning environment that teaches AI agents to pause and confirm before an irreversible action, built on Indian digital public infrastructure.

• Top 9 of 70,000+ registrations at the Meta x Scaler OpenEnv hackathon. Graded by a 6-component deterministic reward with no LLM-as-judge, so every score is auditable, across 43 scenarios and four difficulty tiers.
• Baselines flat-lined at 0.02/1.0 until I traced it to a multiplicative rubric that compounded every small penalty, and redesigned it into a mixed form. Trained Qwen-1.5B via GRPO to 45% of Claude Sonnet's hardest-tier score at 0.05% of the parameters.
• The headline finding was a failure, and I led with it: the 1B model trained to look competent while firing 5 of 5 must-not-execute safety traps.

Python · PyTorch · GRPO / TRL · Unsloth Live demo GitHub

gravel

An embeddable key-value storage engine in Go, a full LSM-tree written from scratch because I wanted to know how Cassandra and RocksDB actually work.

• Hand-rolled every part: skip-list memtable, write-ahead log with CRC32 and crash recovery, SSTables with a bloom filter over xxhash double-hashing, a B+tree block index, and leveled compaction.
• The hard part was correctness under failure, so it has fault-injection tests that crash the engine mid-write plus concurrent stress under the race detector. Benchmarked at ~295K sequential writes/sec, roughly 45% faster than BadgerDB.

Go · LSM-tree · WAL GitHub

OnCall-Env

A "SWE-bench for incident response": the agent triages an alert, investigates, finds root cause, remediates, and writes it up.

• 48 scenarios across 11 real root-cause types (connection-pool exhaustion, memory leaks, replication lag, deadlocks, DNS failure, TLS expiry), under partial observability so the agent has to investigate rather than guess.
• I hardened it over six measured rounds, re-running a frontier model each time to prove it was getting harder in the right way. The gates cut frontier zero-shot success from 0.89 to 0.73.

Python · FastAPI · Docker · TRL GitHub

Agent systems, built at hackathons

Multi-agent products where the interesting part is what the agent refuses to do.

• Vaada (Sarvam Epoch Buildathon): a Telugu voice agent that inherits a departed insurance agent's orphaned policyholders and carries a promise ledger of prior commitments. It will not confirm a promise that is not in the ledger. The refusal is the product.
• SafeHire: a LangGraph auditor for Indian hiring compliance, five agent roles with a human-in-the-loop gate, three deterministic validators against fabricated citations, and state checkpointed so a run resumes after a pause.
• Lumen: an insurance subrogation officer built as a courtroom, where opposing counsel argue issue by issue and a citation gate rejects any claim without real evidence.

Python · LangGraph · FastAPI · Next.js GitHub

Yugaantar

Chairperson, SST's first student-led fest

Ran an ₹55L budget with 21 HODs and 200+ organizers, after two earlier attempts never got off the ground.

• Drove 6,500+ registrations from 800+ colleges and nine podium finishes against teams from the IITs, NITs, and BITS. Covered in the Times of India.
• Rebuilt the sponsor pipeline from scratch after every sponsor withdrew, and swapped the proshow vendor three days out when they failed diligence.

Student Council

First President, Scaler School of Technology

First person to hold the role, so there was nothing to inherit. Work ran on individual effort and things quietly got missed.

• Rebuilt how the council operated: distributed ownership and written workflows instead of one person carrying a department, so nothing stalled when someone was unavailable.
• Created the inter-house system around the college's RICE principles (Respect, Integrity, Curiosity, Excellence) to keep competition alive between events rather than only during them.
• Ran a full calendar of events and gathered feedback from students, clubs, and staff to find where execution was actually breaking, then fixed the process rather than the symptom.

BITS Pilani

Undergraduate degree in Computer Science (Honours)

Studying at Scaler School of Technology in Bengaluru, with BITS Pilani as the degree partner.