About
I have spent 15+ years in the software world, and in the last few years my focus has shifted fully into Artificial Intelligence and Machine Learning. Today, as an AI/ML Architect at Ignite AI, I design and build intelligent systems that turn ambitious ideas into real, production-grade solutions. At Ignite AI I work end to end across architecture, research, and implementation. Some of the things I have been building recently: 1. DynaTune Lite and DynaTune Enterprise – a no-code LLM fine-tuning framework that automates model selection, LoRA/QLoRA/full fine-tuning strategy, GPU sizing and cost estimation, with both "Auto" and "Expert" modes for teams. 2. A Mixture of Recursions (MOR) model, inspired by modern small-model architectures, to get GPT-like behavior in compact, efficient models that are cheaper to serve. 3. A hybrid AI search stack that fixes the limitations of vanilla Elastic based keyword search, combining vector search, query understanding and ranking to handle noisy, real user queries. 4. A generic time-series recommendation and anomaly detection engine that can plug into ARIMA, NeuralProphet, Google TimesFM and similar models to surface trends, forecasts and anomalies across any tabular business data. I also actively build and share in the open source ecosystem. Recently I released: opd-synthetic-distillation – an on-policy distillation framework that combines GOLD-style on-policy learning with synthetic data generation (OpenAI or NeMo) to distill large LLMs into smaller student models while controlling quality and cost: https://github.com/debankadas/opd-synthetic-distillation Before Ignite AI, I shaped AI product strategy as an AI Product Owner at Amdocs, and earlier I was a Tech Lead at Tech Mahindra leading large-scale telecom transformation programs. Those roles taught me how to balance research with delivery, align stakeholders, and ship AI systems that work in production, not just in notebooks. I thrive at the intersection of deep technical work and real-world business impact: designing LLM and ML platforms, distillation pipelines, RAG and search systems, and helping teams move from "we should use AI" to concrete, measurable outcomes. You can read some of my writing here: Medium – https://medium.com/@debanka-das Always open to conversations about LLM training and distillation, AI platforms, and building efficient, real-world AI systems.
Work experience
AI/ML - Solution Architect at Ignite AI, Jul 2025 - Jul 2026 · 1 yr 1 mo, Pune District, Maharashtra, India, At Ignite AI, I am working on scalable data solutions powered by AI/ML, focused on improving data quality, completeness, and trust across diverse sources while maintaining both performance and flexibility. The first product I worked on is Dynatune, a fine-tuning framework that allows anyone to bring their data in any format and fine-tune any model of their choice without needing prior knowledge of AI or ML concepts or parameter values. At the same time, data scientists can use Dynatune with full control, setting and adjusting all parameters as needed. Dynatune - a unified framework that handles everything from training to production inference to eval and benchmark, all while slashing costs upto 80%.
Technical Product Owner - AI at Amdocs, Jan 2023 - May 2025 · 2 yrs 5 mos, Pune District, Maharashtra, India
Tech Lead at Tech Mahindra, Oct 2021 - Jan 2023 · 1 yr 4 mos, Pune Division, Maharashtra, India
Software Engineering Specialist at Amdocs, Jun 2019 - Sep 2021 · 2 yrs 4 mos, Pune District, Maharashtra, India
Software Development Specialist at Amdocs, May 2017 - Oct 2019 · 2 yrs 6 mos, Kuala Lumpur, Federal Territory of Kuala Lumpur, Malaysia, CRM solutions BHD Malaysia
Senior Subject Matter Expert at Amdocs, 2010 - Apr 2017 · 7 yrs 4 mos
