LLMs & Agents
AI assistants that answer customers, search company knowledge, and query databases in plain English, plus small in-house models that cut API bills and keep data private.
See 3 case studies →Principal Data Scientist · ML Engineer · Pune, India
I build production ML systems that move real business numbers.
For ten years I've turned AI ideas into products that ship and pay off: AI assistants that resolve real customer issues, AI video fast enough to sell, and ad pricing that runs on 50B+ requests a day. I lead a data science team at Talentica Software and advise companies on making AI work.
Now · Oct 2026 Leading a five-person team on AI agents and generative video. Open to Principal Data Scientist roles and select advisory work.
- 10 yrs
- building ML & software
- 50B+
- bid requests/day served by my pricing engine
- 500K
- users/month on an AI support assistant
- 70%
- of conversations resolved without escalation
- 47%
- faster AI video generation
- 2×
- cheaper AI serving with in-house models
About
Research depth, production discipline.
I'm a Principal Data Scientist with an M.Tech in Computer Science (Machine Learning) from the Indian Institute of Science, Bangalore. Over the last decade I've moved from backend engineering to leading ML end to end: framing the problem, picking the right model family, and owning latency, cost, and the metric the business actually cares about.
Today I lead a team of five data scientists building AI agents, small in-house language models, AI video generation, and real-time ad pricing. My work is deliberately broad: simple, proven models where they do the job, and frontier generative AI where nothing else will.
I want to keep growing as a Principal Data Scientist: setting technical direction for ML organisations, mentoring the next generation of scientists, and helping companies make sound, measurable AI bets. That's also why I take on selected advisory work.
- LLM fine-tuning & distillation
- Multi-agent systems
- RAG & NL2SQL
- Diffusion video
- Real-time bidding
- Experimentation & A/B testing
Expertise
Four areas I've shipped in production
Pick an area to jump to the case studies behind it.
Generative Video & Vision
AI video that animates a face or body from a single photo, and real-time vision that reads documents and tracks objects in live video.
See 3 case studies →Ad Tech & Classical ML
Pricing and filtering decisions made in a millisecond at massive scale, plus forecasting and anomaly detection, all proven with live A/B tests.
See 3 case studies →ML Platforms & Serving
The engineering that makes models fast, affordable, and reliable in production: multi-GPU serving, monitoring, and automatic retraining. It sits underneath all of the above.
See all case studies →Selected work
Case studies
Each one started as a business problem, not a model. Open any tile for the challenge, approach, and results.
Ad Tech · Classical ML
Real-time bidding: bid filtering & floor-price optimization
Smarter ad pricing that earns more per ad while dropping traffic that was never going to pay.
- 10%+ revenue uplift
- 40% low-value traffic filtered
- 30% infra cost cut
Challenge Price ad inventory well and cut wasteful traffic across 50B+ bid requests a day from 100+ bidders, inside a hard latency budget.
Approach & outcome
- Architected a real-time floor-pricing engine: a multi-armed bandit layer on top of CatBoost base-price regression.
- Built a CatBoost response-probability filter at 1 ms p99, with a dynamic threshold that caps revenue loss at ~1%.
- Validated through closed-loop A/B testing with continuous recalibration, for a sustained 10%+ revenue uplift.
CatBoost · Multi-armed bandits · A/B testing
Generative AI · Agents
Multi-agent AI system for support automation
An AI support assistant that resolves most customer conversations without a human.
- 500K users / month
- 70% resolved unescalated
Challenge Automate high-volume customer support without losing the safety net of human escalation.
Approach & outcome
- Architected a production multi-agent system on LangGraph: an intent router delegating to FAQ, recommendation, and ticketing agents.
- Built structured memory, state management, and failure recovery for long-running conversations.
- Instrumented evaluation and observability with LangFuse so quality is measured, not assumed.
LangGraph · LangChain · LangFuse
Generative AI · SLM fine-tuning
Resume intelligence: CV parsing & JD relevance scoring
Replaced an expensive external AI service with small in-house models: candidate data stays private, at half the cost.
- 2× lower serving cost
- 86 / 83% parsing P / R
- 84 / 88% scoring P / R
Challenge Replace a frontier-LLM system with self-hosted models to keep candidate data private, while holding quality and cutting cost.
Approach & outcome
- Distilled two small language models (QLoRA/PEFT) from human-validated frontier-LLM labels.
- Parsing: fine-tuned Qwen3-0.6B to extract structured CV fields.
- Relevance: fine-tuned Qwen3-1.7B to score JD–resume fit with an explanation.
- Served with vLLM on AWS at half the cost of the previous system.
Qwen3 · QLoRA · PEFT · vLLM · AWS
Generative Video · Emovid
Expression & pose transfer: talking-head animation
Turns a single photo into a realistic talking video, almost twice as fast as before.
- 47% lower per-clip latency
- 170s → 90s per clip
- 10B+ parameter model on 4 GPUs
Challenge Turn a single image into a realistic talking video, fast enough to be a product rather than a demo.
Approach & outcome
- Architected an end-to-end GPU pipeline that pairs a landmark detection model with a diffusion model conditioned on body pose and facial features for expression transfer.
- Earlier, trained a talking-head model from scratch with multi-GPU training on 8 GPUs over ~10 hours of video.
- Deployed a 10B+ parameter model sharded across 4 GPUs; generating at half frame rate and interpolating the missing frames cut latency 47%.
Diffusion models · Landmark detection · Multi-GPU inference · Frame interpolation
Generative AI · RAG
Multimodal RAG & NL2SQL enterprise assistant
Lets business users ask questions of their databases and documents in plain English.
- 87% SQL execution accuracy
- 70% less manual SQL
- 92% recall@15 over ~1M chunks
Challenge Let business users query structured data in plain language and search text, image, video, and audio knowledge in one place.
Approach & outcome
- NL2SQL with schema linking, query validation, and self-correction loops.
- Multimodal RAG over text, images, video, and transcribed audio.
- Dense retrieval with metadata filtering and BGE-reranker-v2 on PGVector.
PGVector · BGE-reranker-v2 · PostgreSQL · LLMs
Computer Vision · Privacy
Real-time document alignment & ID masking
Straightens scanned ID documents and hides sensitive numbers automatically, in under a fifth of a second.
- 99.5% recall on ID fields
- 4° orientation RMSE
- <200 ms end to end
Challenge Correct document orientation and mask sensitive IDs in real time, where a single miss means PII leakage.
Approach & outcome
- Designed a YOLOv8 + MobileNet + OpenCV pipeline for orientation correction and identity-field detection.
- Deliberately tuned for recall over precision to minimise PII exposure.
- Shipped as a production API under 200 ms end to end.
YOLOv8 · MobileNet · OpenCV
Classical ML · MLOps
Automated forecasting pipeline with drift-triggered retraining
A forecasting system that notices when its data changes and retrains itself.
- 0.90 R²
- Auto retraining on drift
Challenge Forecast performance from geographic, census, and demographic signals, and keep the model accurate as the underlying data shifts.
Approach & outcome
- Built an end-to-end pipeline on GCP: data in BigQuery, model training and deployment with AutoML.
- Joined geographic, census, demographic, and historical data into a single feature set.
- Added KS-test drift monitoring on incoming data that triggers automated retraining, with no manual babysitting.
GCP AutoML · BigQuery · KS-test drift monitoring
Computer Vision · Video
Real-time multi-object tracking
Follows people and objects through live video while keeping track of which is which.
- 20 FPS on video streams
- Persistent identities
Challenge Detect and follow multiple objects across live video, keeping each object's identity consistent from frame to frame.
Approach & outcome
- Used YOLO for per-frame object detection.
- Paired it with DeepSORT for identity-persistent tracking across frames.
- Ran the combined pipeline at 20 FPS across video streams.
YOLO · DeepSORT · OpenCV
Citrix · Anomaly Detection
Anomaly detection on server telemetry
Flags real server problems while cutting false alarms by almost a third.
- 30% fewer false positives
Challenge Flag genuinely abnormal server behaviour in Citrix Analytics without flooding operators with false alarms.
Approach & outcome
- Built Isolation Forest models on server response telemetry.
- Brought false positives down by 30%, so alerts were worth acting on.
Isolation Forest · Python
Advisory
Bring in a Principal-level ML lead, without the full-time hire
I work with founders, CTOs, and data teams who need a senior ML voice to get a decision right or a system into production. Every engagement is scoped around a measurable outcome.
What I can take on
Strategy & feasibility
Is this an LLM problem, a classical ML problem, or not an ML problem at all? A clear build/buy/skip call with cost and accuracy estimates.
Architecture & build
Hands-on design and delivery, with evaluation and observability built in from day one rather than bolted on after launch.
Cost & latency reduction
Distilling frontier APIs into self-hosted models, multi-GPU inference, and serving optimizations for systems already in production.
Reviews & mentoring
Architecture reviews, experimentation design, and coaching so your team can run the system confidently after I step back.
How an engagement runs
- 1
Diagnose
Business goal, data, constraints, and the metric that defines success.
- 2
Prototype
A fast baseline against an honest evaluation set before anyone scales anything.
- 3
Productionize
Latency, cost, monitoring, and A/B validation in the real environment.
- 4
Hand over
Documentation, dashboards, and a team that owns it.
“We have absolutely loved working with Alakh and he has been a cornerstone of our team, thanks so much!”
-
Talentica Software
Oct 2019 – Present · Pune
Lead a team of five data scientists delivering production ML across LLM agents, generative video, and ad tech. See the case studies.
- Principal Software Engineer – Data ScienceOct 2023 – Present
- Senior Software Engineer 2 – Data ScienceOct 2020 – Oct 2023
- Senior Software Engineer 1 – Data ScienceMar 2020 – Oct 2020
- Software Engineer 2 – Data ScienceOct 2019 – Mar 2020
These are Talentica's official titles; its data science track uses "Software Engineer – Data Science". My current role is the equivalent of Principal Data Scientist.
-
Citrix R&D India · Software Engineer 2
Jul 2018 – Oct 2019 · Bangalore
Anomaly detection for Citrix Analytics (30% fewer false positives) and NetScaler features.
-
Soft Solutions · Software Engineer
Jul 2013 – Jul 2015 · Jaipur
Java and MySQL backends, deployed on-premise for small businesses.
Education
-
M.Tech, Computer Science & Automation (ML) · Indian Institute of Science, Bangalore
2016 – 2018
Research on multi-class imbalance and debiasing word embeddings. AIR 45, GATE 2016 (CS).
-
B.Tech, Computer Science Engineering · Rajasthan Technical University
2009 – 2013
Toolkit
- LLMs & agents
- QLoRA/PEFT, distillation, RAG, LangGraph, LangChain, LangFuse · familiar with RLHF, PPO, DPO, GRPO
- ML & stats
- CatBoost, XGBoost, multi-armed bandits, A/B testing, survival analysis
- Deep learning & CV
- PyTorch, Transformers, diffusion models, YOLO, OpenCV
- Distributed & serving
- DDP, FSDP, multi-GPU inference, vLLM, ONNX
- Cloud & data
- AWS, GCP (AutoML, BigQuery), RunPod, Docker, PySpark, PostgreSQL, PGVector
- Languages
- Python, SQL, Java
Writing
Published articles
- KDnuggetsAug 2026 Beyond Bots: Rethinking AI Support with a Hybrid AI Architecture Why dependable support assistants blend retrieval (RAG) with fine-tuning to stay safe, accurate, and fast.
- Analytics InsightJun 2025 Optimizing AI Performance: Contextual Input vs. Fine-Tuning in LLMs When to give an LLM context in the prompt and when to fine-tune it, weighed on accuracy, cost, and freshness.
- Talentica BlogOct 2023 AI for Recruiting: How AI Can Help Recruiters Boost Productivity Recruiters spend about 40% of their time finding candidates; how AI can take that load off.
- VentureBeatJan 2023 Adopting AutoML: Let's Do a Reality Check What AutoML can and can’t do for a team, before you commit to it.
- KDnuggetsMay 2022 Operationalizing Machine Learning from PoC to Production ML only pays off in production; how to get a project past the proof-of-concept stage.
Contact
Have a hard ML problem? Let's talk.
For principal or leadership roles in data science, advisory engagements, or speaking and writing, send me a message here and it comes straight to my inbox.
- Emailalakhsharmacs@gmail.com
- LinkedInlinkedin.com/in/alakhsharma
- GitHubgithub.com/alakhsharmaa
- Based inPune, India · working with teams worldwide