arXiv:2606.00001v1 [cs.SE] · Bengaluru, India · 2026
Building AI Systems That Survive Production:
A Portfolio
AI Systems Engineer, IBM Software Labs · shreyash.pandey.katni@gmail.com
Abstract
This portfolio documents work at the intersection of applied AI, backend engineering, and reliability: a 125M language model trained from scratch on a single consumer GPU, multilingual Indic models, agentic publishing and security systems, and a three-tier locator auto-healing system shipped inside IBM's test automation. We report honest benchmarks and provide four interactive artifacts that run entirely in the reader's browser.
1Selected results
The systems below were evaluated under internal or public benchmarks. Numbers are reported as measured, including where they are near chance.
| Metric | Value |
|---|---|
| test locators auto-healed | 5K+ |
| availability supported | 99.99% |
| fewer support tickets | 70% |
| TechInterrupt Hackathon, India | 1st |
| Chrome fallback accuracy | 83% |
| tokens trained from scratch | 2B |
2Systems
2.1 Phoenix 125M (Released)
A LLaMA-style 125M decoder-only model trained from scratch on a single RTX 3080 Ti with a custom tokenizer, data pipeline, and training loop.
| Property | Value |
|---|---|
| Parameters | 125M |
| Tokens | ~2B |
| WinoGrande | 0.507 |
| License | Apache 2.0 |
2.2 Sweta-Hi / Sweta-Kn (Released)
Hindi and Kannada pretraining on a LLaMA-style architecture with a joint 64K tokenizer and an end-to-end multilingual data pipeline.
| Property | Value |
|---|---|
| Parameters | 134M |
| Vocab | 64K |
| Hindi PPL | 14.5 |
| Kannada PPL | 34.0 |
2.3 LinkedIn Post Swarm (Live)
A multi-agent publishing workflow: a research agent, a Claude orchestrator, a critic loop, and a Playwright publisher with a Telegram human-in-the-loop.
| Property | Value |
|---|---|
| Engagement | 5x in 1mo |
| Follow reqs | ~1/day |
| Agents | 4 |
2.4 Rudra (Parked)
A multi-agent offensive-security architecture with hard-coded scope guardrails, sandboxed execution, and auditable event-driven workflows.
| Property | Value |
|---|---|
| Agent types | 4 |
| Implemented | 2 / 4 |
| Cluster VRAM | 22 GB |
2.5 LocalLeads (Live)
End-to-end local-business discovery, AI site generation, deployment, and personalized outreach. A multi-hour agency process compressed into minutes.
| Property | Value |
|---|---|
| Site gen | < 2 min |
| Cost/site | Rs 500 |
| Contacted | ~25 |
3Interactive artifacts
All four artifacts execute client-side. No inference server is contacted; computation runs in JavaScript on the reader's device.
Figure 1. BPE tokenization under the Phoenix 32K English vocabulary versus the Sweta 64K joint vocabulary. Note the fragmentation of Indic scripts under the English-only tokenizer.
Figure 2. Semantic retrieval over the project corpus via TF-IDF and cosine similarity, computed in-browser.
click( locator: "#login-submit" )Figure 3. The three-tier locator auto-healing ladder. Recovery escalates from exact CSS match to text embedding to a vision-language model.
Figure 4. Commonsense reasoning game versus Phoenix-125M. The model scores ~0.507 on WinoGrande (chance = 0.50).
4Experience
Software Development Engineer 2, IBM Software Labs (2024 - Present). Reliability-heavy AI capabilities inside browser automation and testing products: embeddings, vision-language models, service decomposition.
Software Engineer 2, Software AG (now IBM) (2023 - 2024). Semantic retrieval, internal copilots, and prediction systems. Won the TechInterrupt Hackathon, first in India.
Software Engineer, Software AG (2022 - 2023). Enterprise integration platform work across Java, Spring Boot, and REST APIs. 2023 Star Performer.
References
- Shreyash Pandey. HuggingFace model releases. https://huggingface.co/shreyash-Pandey-Katni
- Shreyash Pandey. Source repositories. https://github.com/shreyash-Pandey-Katni