About me
Principal AI Engineer. Six years of getting models past the demo.
Speech, documents, LLMs, and all the unglamorous infrastructure that decides whether any of it survives contact with production.
I build AI systems that hold up outside a notebook. Most of my work sits in three places: speech intelligence, LLM automation, and document processing, usually all three at once, because that's what an enterprise problem actually looks like.
The last few years have been spent architecting AI platforms for enterprise clients: real-time voice agents, document intelligence suites that chew through unstandardized PDFs, multi-tenant systems where one client's data must never touch another's. I don't just train the model. I build the queue that retries it, the schema that isolates it, the observability that tells you when it quietly starts getting worse.
I like the messy end of the problem. Give me an idea that three people describe three different ways, and I will turn it into something a team can actually run, extend and trust. I'm opinionated about modular design, honest metrics, and never shipping a system nobody can debug at 2am.
Away from the keyboard, I'm usually mentoring engineers or working out how to make GPU workloads cost less than they did last month. The open-source work on this site is the other outlet: if I learn something worth knowing, I'd rather write it down than let it evaporate.
Experience
One company, three jobs, six years
I joined TrellisSoft in Bengaluru straight out of college and grew with the AI practice there, from training the first models to owning how the platforms are built and run.
July 2025 to present
Principal AI Engineer, TrellisSoft
Architecture and delivery of the AI platforms across speech, LLM and RAG work: high-throughput inference on vLLM and partitioned GPUs, multi-service orchestration, CI/CD for model workloads, and the conversations with leadership about what the AI roadmap should actually be.
June 2023 to June 2025
Senior AI Engineer, TrellisSoft
Led the multilingual speech intelligence, real-time agent assist and call QA platforms; built RAG knowledge systems and layout-aware OCR pipelines for insurance and finance; fine-tuned and served domain LLMs (Qwen, SmolLM, OLMo) on vLLM. Mentored the team and set the design patterns it still uses.
August 2020 to June 2023
AI Engineer, TrellisSoft
NLP, speech recognition and computer vision models in production: chatbots, sentiment analysis, document recognition, with the semi-supervised data pipelines and retraining workflows that kept them current. Brainiac of the Quarter in 2021, Employee of the Quarter in 2022 and 2023.
2016 to 2020
B.E. in Computer Science, MVJ College of Engineering
Bengaluru, with an 8.0 GPA.
How I work
The shape of a project, most of the time
Not a methodology with a logo. Just the order things tend to happen in when the result has to survive real users.
First
Find the failure mode before the architecture
Every AI product has one thing that decides whether it works: the accent the ASR misses, the scanned page the OCR mangles, the query the retriever answers confidently and wrongly. Everything else gets easier once that's named.
Then
Prototype ugly, harden deliberately
The first version exists to be measured, not admired. Once the numbers say the approach holds, it gets rebuilt properly: boundaries, retries, idempotency, and a schema that won't need a migration in a month.
Always
Make the system explain itself
Models degrade quietly. Logs, traces and per-tenant metrics aren't a nice-to-have on an AI platform; they are the only way anyone finds out that last week's fine-tune made things worse.
Finally
Hand it over so it outlives me
Documentation, runbooks and a team that understands the trade-offs. A platform only one person can operate isn't finished, however good the metrics look.
Toolkit
What I reach for
Grouped honestly: the things I use every week, not everything I've ever opened a tutorial on.
Languages & core
- Python
- SQL
- Bash
- JavaScript / React
Models & training
- PyTorch
- Hugging Face Transformers
- LoRA / QLoRA fine-tuning
- Unsloth
- LangChain
- SentenceTransformers
- scikit-learn
Speech & documents
- FasterWhisper / WhisperX
- pyannote.audio
- Kaldi
- TTS: Kokoro, XTTS, Bark
- Surya OCR / PaddleOCR
- PDFPlumber
Serving & infrastructure
- vLLM
- llama.cpp / Ollama
- Triton Inference Server
- FastAPI / Flask
- Docker & Compose
- NVIDIA MIG
- Celery / Redis queues
- Nginx
- AWS / Azure / on-prem GPUs
- GitHub Actions / Jenkins
Data, security & observability
- PostgreSQL / TimescaleDB
- Redis
- Qdrant / FAISS / Chroma / pgvector
- Multi-tenant design
- Vault, JWT, 2FA, RBAC
- Grafana / Loki / OpenSearch
Say hello
Let's talk
Whether it's a role, a system that needs rescuing, or a question about something on this site, email is the fastest way to reach me.