M.Tech, Information Technology
Netaji Subhas University of Technology
Advanced DSA · Machine Learning · Deep Learning · NLP · Blockchain
Sayantan Sikdar
AI/ML Engineer & IEEE-published researcher — building LLM systems, durable agents and data pipelines that hold up in production.
Scroll to awaken
(00) About魂
I'm an M.Tech candidate in New Delhi who ships end-to-end AI systems — fine-tuned LLMs, durable agents, a lakehouse reconciled to the cent, and software verified by hundreds of tests. Published at IEEE. I own the loop from scoping with stakeholders to explaining the numbers.
- Years in industry
- 0+
- Projects shipped
- 0
- IEEE publication
- 0
- Certifications
- 0+
Concentrate.
Keep scrolling.
Bankai.
To master the blade
is to master the self…
to master the model
is to master the data.
(01) Arsenal力
The arsenal
Seven disciplines, one blade. The tools I reach for when a problem needs to be measured, modelled and shipped.
-
一
Languages
-
二
LLMs & GenAI
-
三
Machine learning
-
四
Data engineering
-
五
Backend & orchestration
-
六
Testing & QA
-
七
Cloud & DevOps
(02) Selected work戦
Selected work(09)
-
01
$295.6KReconciled
NovaCart Medallion Pipeline
Microsoft Fabric lakehouse for a 5-country retailer whose finance team saw conflicting revenue. Guarded Delta MERGEs, SCD Type 2 and as-of FX turned 2,963 order versions into 945 orders, blocked 10 stale updates, and matched every gold number to an independent check — 224 tests, 0 rows changed on re-run.
-
02
529Tests
SmartDialer
Outbound-dialer simulator where every dial passes 13 explainable safety rules. 529 Vitest tests and 35 Cypress–Cucumber scenarios gate CI with 0 do-not-call breaches; a covering index made reservation 40× faster at 1,500 agents.
-
03
1 : 1Workflow / order
Order Supervisor
Long-running AI agent on Temporal: one durable workflow per order wakes on 3 triggers, a classifier gates events by urgency, and the agent acts through 6 tools — sleeping on durable timers instead of polling.
-
04
99.4%Params cut
EmpaChat — Empathetic AI
GPT-2 XL fine-tuned with LoRA and 4-bit QLoRA on 58K GoEmotions comments: 8.4M of 1.56B parameters trained for ~89% emotion accuracy at ~150 ms GPU latency, served via FastAPI and Gradio.
-
05
3.851Sharpe
Basket Trading Optimisation
Benchmarked 6 optimisers for cointegration basket weights out of sample. CVFS-CMA-ES hit a 3.851 Sharpe — 9× standard CMA-ES — with −3.99% max drawdown; Bayesian optimisation returned 63.76%.
-
06
99%Accuracy
Fruit Grading with Deep Learning
AlexNet and ResNet v2 framework for real-time, non-destructive grading — 99% accuracy on 2.2K images and 98.33% (0.9752 F1) on 15.5K. Read the IEEE paper ↗
-
07
93%mAP
YOLO Automatic Meter Reading
Built at WBSEDCL: YOLO v5/v7/v9-e digit recognition with erosion and HSV correction on 466 images — 99% accuracy, 95% fewer errors, 500+ staff-hours saved a year.
-
08
40hBuild time
Morsaab’s Restaurant Platform
Freelance: Next.js 15 platform with a 9-category menu on Neon Postgres, Stripe Checkout, a Clerk-gated admin CRM and Datadog RUM — 9.8k lines in under 40 hours, live on Vercel.
-
09
6Services
Bookie — MLOps Recommender
Full recommender lifecycle in one compose file: a 4-stage Airflow DAG trains SVD collaborative filtering, with DVC, MLflow, FastAPI, React and Prometheus/Grafana monitoring.
(03) Path道
The path
-
Jul 2025
AI Evaluation Specialist
Handshake
- Contributed 6+ Docker-sandboxed agentic-AI evaluation tasks to Terminal-Bench 2 (Harbor) for frontier-model benchmarking.
- Enforced four-way consistency and tamper-proof verifier design — 88.9% acceptance rate, meeting the GPT-5.4 acceptance bar.
-
Feb 2025 — May 2025
Freelance Full Stack Developer
Self-employed · Morsaab’s
- Built a Next.js 15 + TypeScript restaurant platform: database-driven menu, Stripe Checkout, Clerk-gated admin CRM and Datadog RUM.
- Shipped 9.8k lines in under 40 hours — about 4× faster than a typical 3–4 week build; live on Vercel.
-
Mar 2024 — Apr 2024
Architect-Machine Learning Trainee
WBSEDCL
- Architected a YOLO v5 / v7 / v9-e automatic meter-reading system — 99% accuracy, 94% precision, 93% mAP.
- Erosion and HSV correction on 466 images cut errors by 95%, saving a projected 500+ hours a year.
-
Sep 2023 — Mar 2024
Artificial Intelligence Intern
NIELIT Kolkata
- Designed a DNN grading framework with AlexNet and ResNet v2 — the work was published at IEEE.
- 99% accuracy on 2.2K images; 98.33% with 0.9752 F1 on 15.5K images.
(04) Training grounds学
B.Tech, CSE (AI & ML)
Haldia Institute of Technology
Time-series analysis · Probability & statistics · Software engineering · DSA · DBMS · OS
- C—01Google Cloud Training — Cloud Engineering Fundamentals & Essentials, Cloud Application Development, Cloud ML-AI, DevOps & Kubernetes
- C—02Anthropic — Claude 101, Claude Code 101
- C—03NVIDIA — Fundamentals of Deep Learning
- C—04AWS Machine Learning Scholarship 2021, Udacity — Recipient
- C—05GATE CS 2024 — Qualified
- C—06Youngest IEEE Student Researcher
(05) Research書
IEEE
International Conference · 2024
Deep Neural Network Based Yield Phenotyping Framework
A deep neural framework built on AlexNet and ResNet-v2 for automated fruit grading — 99% accuracy on 2.2K images and 98.33% accuracy with a 0.9752 F1-score on 15.5K images.
Read on IEEE Xplore(06) Contact絆
Let's build something intelligent.
sayantansikdar@outlook.com
- Phone
- +91 89020 91002
- Based in
- New Delhi, India