ai engineer · bengaluru

building smart agentsand intelligent systems

the training manifold

layer ℓ = 15 / 32

1 subtle5 noticeable20 things get weird

α 0.0· autoplay

prompt

who is kishan?

output

α too large → off the training manifold → hallucination. read more

01 / about

who's building this

i'm an ai engineer who spends most days building agentic systems: multi-agent pipelines, retrieval stacks, and the unglamorous plumbing that makes an llm reliable in production.

outside of work, i've spent the past year-plus reading and writing about mechanistic interpretability, trying to understand not just what these models do, but how they do it. steering vectors, sparse autoencoders, the geometry of the residual stream.

i also build solara, a video tool that turns raw screen recordings into polished demos, on nights and weekends. i'd rather ship a sharp, focused tool than a flexible one that tries to be everything.

currently

building solara
ai engineer at valtech
reading & writing on mech interp
based in bengaluru, india

02 / work

experience

ai engineer · valtech
may 2026 - present
bengaluru, india · on-site
- building an ai platform that automates sdlc documentation and review workflows, cutting turnaround time and standardizing delivery artifacts across engineering teams.
- building a meta harness for modernizing legacy .net systems of 1m+ loc with a multi-agent architecture.
1m+ locmulti-agent
ai engineer · nds
july 2025 - april 2026
navi mumbai, india · on-site
- built zara, an ai assistant (azure openai, cosmos db, azure ai search) used by 3000+ employees for day-to-day hr and it support.
- automated freshservice ticket categorization, cutting it workload by 30%, and built a sap successfactors leave-management agent that reduced hr workload.
3000+ employees−30% it workload
ai-ml engineer intern · wasserstoff
november 2024 - june 2025
gurugram, india · on-site
- developed llm fine-tuning workflows: supervised fine-tuning and reinforcement learning (grpo) on llama 3.1 8b and qwen 2.5 7b, on custom indian law datasets.
- built a large-scale document processing pipeline for litlaw's rag, handling 100k+ pdfs with ocr, summarization, keyword extraction, and embeddings in qdrant.
- built chatur ai, a multi-agent hr automation system (fastapi, mongodb, langgraph, celery) that cut hr workload by 50%+.
100k+ pdfs50%+ hr workload cut
data science intern · wise analytica
may 2023 - february 2024
london, england · remote
- used the twitter api with ner and ml-based lead scoring to identify prospective customers, a 60% increase in lead generation.
- built an nlp model matching occupation titles to degree titles at 85% accuracy.
+60% leads85% accuracy

03 / projects

things i've built

featured

solara

transforms raw screen recordings into polished, demo-ready videos within minutes, using ai-generated voiceovers, intelligent zoom effects, and script optimization, built for training material, presentations, and marketing.

“after quite a few weeks of building, iterating, and obsessing over every frame, i'm finally launching solara.”

pythonffmpegopencv

visit solaraai.in

catch-cap

llm hallucination detection library. multi-signal detection combining semantic entropy, token log-probabilities, web-grounding, and llm-judge verification into a unified 0–1 confidence score. supports openai, gemini, and groq with mixed-provider configs and graceful degradation.

pythonopenaigeminigroqtavily

litlaw pipeline

large-scale document processing pipeline for rag over indian law: 100k+ pdfs, ocr, summarization, keyword extraction, embeddings in qdrant, metadata in mongodb.

ragqdrantmongodbllama 3.1

chatur ai

multi-agent hr process automation system for screening, scheduling, and coordination, cutting hiring time and effort drastically with minimal manual oversight.

fastapimongodblanggraphcelery

04 / writing