sam42

Two tracks. One consultant.

I specialize in AI/ML systems and high-performance distributed infrastructure. Both tracks are deep enough to be full-time disciplines : I operate across both, which is where the hard problems live.

Track 1 : AI/ML Systems

From custom neural network architectures to production LLM pipelines, I help teams build AI systems that actually work in the real world : not just in demos.

Neural network design & training

  • Design and train custom CNN, GAN, and hybrid architectures on proprietary datasets
  • Frameworks: PyTorch, Theano, MXNet, CUDA
  • Specialization: generative models (StyleGAN), computer vision, time-series
  • Outcome: production-ready models deployed at scale

LLM systems & agentic pipelines

  • RAG pipelines with hybrid vector search and knowledge graphs
  • Autonomous agents for decision-making, monitoring, and alerts
  • DSPy for modular, maintainable AI pipelines
  • LLM-as-Judge evaluation frameworks
  • Outcome: reliable systems beyond prompt engineering

Computer vision & video AI

  • Real-time video analysis in-browser and server-side
  • Attention detection, behavioral analysis, anomaly detection
  • Deploy CV at the edge without crashing

Deep learning consulting & strategy

  • Audit ML systems : bottlenecks, failure modes, scaling
  • Stack, architecture, and training recommendations
  • Parallel distributed training for large models

Time-series processing

  • Designed systems Using both DNN based and Traditional ML based systems.
  • Experience: algorithmic trading, payment intelligence, load prediction

Past contexts: proprietary trading desks, fashion and retail, EdTech, B2B SaaS, fintech.

Track 2 : Low-latency distributed infrastructure

I design and ship infrastructure where latency is mission critical, not a feature. From NSE-certified trading platforms to multi-region video infrastructure : systems at the edge of what's technically possible, kept running in production.

High-Frequency trading systems

  • 2 micro-seconds tick to order latency for NSE at 99.2 percentile on commodity hardware
  • Market connectors, order management engine, risk engines, algo approval documentation
  • Used Kernel bypass for user-space networking along with chache hits optimizations

Multi-region distributed video confrencing SaaS for developing world

  • productise and deployed WebRTC for low bandwidth (>400 kbps) and high availability (6 AWS regions)
  • For real-time experience a client from US connects to a US servers while Indian client would connect to Indian servers
  • In browser classroom with whiteboard, video, presentations and proctoring (vision AI)

Payment orchestration & SOR (Smart Order Routing)

  • Routing engines based on live failure rates, client banking preferences, and gateway capacity
  • Used Cloudflare native tools (workers, KV, Hyperdrive, R2, analytics) along with Supabase (PostgreSQL) for truly serverless architecture
  • Monitoring, alerting, reconciliation and observability

Protein folding simulation process

  • Ported a MATLAB computational biology application to Perl and then to native C
  • Reduced runtime from multiple days to under five minutes drastically increasing experimental bandwidth
  • Key operations: matrix multiplication and large-scale string comparison over protein sequence data

Systems optimization & debugging

  • Latency bottlenecks : kernel and network profiling
  • From “works but slow” to production-grade with benchmarks

Past contexts: proprietary trading desks, EdTech, payment aggregators, global SaaS.

How we work together

Project-based

Define the scope : I deliver the system or component from architecture to deployment.

Advisory

Ongoing consulting on roadmap, architecture, and hiring.

Embedded

Work alongside your team as senior technical capacity for a defined period.

I typically engage with teams who have a hard problem that needs someone who's done it before. If you need a commodity developer, I'm not the right fit.

Let's see if there's a fit →

Who I'm not for

  • Teams looking for a cheap developer or body-shop resource
  • Companies that need AI as a buzzword without a real problem
  • Early-stage ideation without a technical counterpart to work with