I build AI agents that are faster, cheaper and more reliable.
Software Engineer at BrowserStack. I work on AI coding agents, LLM observability and evals. My job is reading traces to find out why an agent was slow or wrong. So this page is one too.
Where the time went
2+ years across backend systems and production AI agents.
BrowserStack
- role
- Software Engineer
- where
- Mumbai
- start
- Sep 2025
- status
- running
- Designed and built a two-language AI Observability SDK (Node.js/TypeScript and Python) that traces LLM apps with zero code changes, using OpenTelemetry across OpenAI, Anthropic, AWS Bedrock, Google Vertex AI and LangChain.
- Built a custom telemetry pipeline: smart span filtering, finding the caller from the call stack, trace propagation across services, and decorators (
@workflow,@agent,@task,@tool) for structured tracing of agent workflows. - Built a scalable LLM evaluation platform with 20+ metrics: faithfulness, answer relevancy, contextual precision/recall, bias detection, safety scoring, RAGAS and G-Eval, with async job processing and concurrent workers.
- Designed a multi-variant experiments framework: dataset versioning, prompt lifecycle management, batch evaluation, human-review annotation queues, and model comparison across quality, latency and cost.
- Improving the production agent's speed, reliability and cost using telemetry.
- Removed an optimisation that was adding model calls. One representative workflow went from 15 → 4 model calls, $0.372 → $0.229 per task and 149s → 36s runtime.
- Tuned prompt-cache use from production telemetry: halved the cache keepalive, saving about $240–300 a month while keeping warm resumes fast. Tested a further cut and rejected it because it hurt resume reliability.
- Found that 58% of production tasks repeated identical tool calls. Uncovered eight separate looping root causes and built safeguards: parallel-safety gates, tool-call batching and completion validation.
- Root-caused a task that read the same file 254 times without finishing: the dedup cache keyed on file path only, not path plus line range. Reproduced it at the wire level and fixed the read tracking.
Credflow
- role
- SDE 1
- where
- Noida
- start
- Aug 2024
- end
- Aug 2025
- Built LLM agents with LangChain and Vertex AI that read complex GSTR-3B CAM sheets (10–15 sub-sheets) to automate credit underwriting, with Kafka for event-driven processing.
- Built backend services for the Loan Origination and Loan Management systems.
The whole run, on one screen
Every bar is a real piece of my work. Click a row to jump to its details.
Things I've built lately
Each one started as a small, real problem. Scroll to replay the session: the problem, the call, the result.
How things really work
Deep dives into everyday engineering, one topic a day, below the surface.
What really happens when you type a domain name
Caches, the recursive resolver, root → TLD → authoritative referrals, glue records, the packet, TTLs and attacks. With real traces.
Sharing a secret without sharing the secret
Diffie–Hellman from paint mixing to clock maths, a live calculator, the man-in-the-middle, X25519 and TLS 1.3.
Let's add one together.
Open to SDE-2 roles in AI agents, LLM tooling and backend systems.