Praneeth Suresh
ML engineer and computer science and mathematics student at the National University of Singapore (double major, 2025–2028). I work on tools that make AI-assisted software development checkable, and I keep my study notes public.
Now
I maintain Beryl, a set of repository files and checks for working with coding agents, and I am researching, as CP3106 independent research at NUS, whether machine-readable repository policies with deterministic enforcement improve coding-agent outcomes over the same rules in prose. The study is in progress and has no results yet.
Experience
AI Legal Automation Intern, Agile Counsel May – Jul 2026
Built a React and Node.js legal document-automation workflow with validated intake and generated incorporation documents.
Co-founder and ML Engineer, Kabal Research Sep 2023 – Jan 2026
Built NLP and graph-analytics pipelines that turned geopolitical news into structured risk signals for analysts.
Co-founder and ML Engineer, ReProse Dec 2023 – Jun 2024
Built an OCR system for handwritten input and a retrieval-augmented generation prototype.
Details, including education and awards, are in the CV.
Work you can inspect
- Beryl — An installable set of repository files and scripts that tell coding agents how to work in a project and check that some of those rules were actually followed.
- Notes — this site — The site you are reading: a static build that turns my Notion study notes into public, searchable pages while keeping math, code, and nested pages intact.
- SIAP — Structure and Interpretation of Agentic Patterns: a book in progress that teaches agent systems by building interpreters for reasoning and tool-use loops.
- AI Society Corpus — An interactive knowledge graph for NUS AI Society that links AI/ML study topics by prerequisites, shared tags, and backlinks.
I also write about research reading and projects and keep computer science notes.
Other public repositories
AI and ML research
LOBForecasting, Sentiment-Analyzer, ReinforcementLearning, CT-image, pv-site-datamodel, ForecastingFastestCoderFirst, Streamlit, MLdrivenDataAnalysisCensus2011, IntroductoryTensorflow, DeepLearningWeek
Software and app systems
Notes, AgentCoding, OpenAIBuild, Praneeth-Suresh, spill, ORCaseAnalysis, Stochron, VentApp, incident-resolving-agent, SnapSteady, skills-github-pages, CyberCupid, reprose, DLW
Energy, data, and experiments
SisThinkers, 2025_aoc, Praneeth-Suresh.github.io, ocf-data-sampler, PVNet
Research questions
The questions I am working on now, why they matter to me, and what evidence would change my view.
Can interpretability become a cheap, standard part of model evaluation instead of a research project per model?
Why it matters to me: My position essay argues that explanation methods are rarely used in real evaluation because they are too costly and not meaningful enough. Probing and representation analysis on frozen models are the most promising candidates because they do not change the model being explained.
What would change my view: Evidence that frozen-model probes routinely disagree with causal interventions on the same model, or that they add nothing over held-out accuracy when choosing between models.
Do machine-readable repository policies, enforced by deterministic checks, make coding agents more reliable than the same rules written in prose?
Why it matters to me: This is the question behind Beryl and my CP3106 independent research. Prose instructions are easy to write but give no evidence they were followed.
What would change my view: If agents given equivalent natural-language guidance pass as many hidden tests and policy checks as agents given enforced policies, the extra machinery is not earning its cost.
Is learning usefully described as finding a low-rank representation of the data, and does that make learned knowledge easier to read out?
Why it matters to me: LoRA and distillation suggest much of what large models store can be re-encoded compactly. If that is general, a compact representation should be convertible into something a person can inspect.
What would change my view: Tasks where adaptation needs high-rank updates to reach full-fine-tuning quality, or where compact representations stay no more interpretable than the full model.
Reading list archive (13 themes)
Deep learning foundations
The first AI research thread is the question behind neural networks themselves: what kinds of functions can trainable systems represent, and what architectural biases make that representation practical?
Architectures for representation learning
CNNs and transformers show the central AI research move: choose a computation pattern that matches the structure of the data, then train the whole pipeline end to end.
Efficient model adaptation
Large models make adaptation a first-class research problem: how much can a model specialize without retraining every parameter or adding unacceptable inference cost?
Controllable generation and diffusion language models
Controllable generation asks whether generation can be steered by explicit conditions while preserving fluency, diversity, and task usefulness.
Memory and test-time learning
Frontier AI systems increasingly need mechanisms for long context, persistent memory, and adaptation during use rather than only during pretraining.
Deep reinforcement learning
Deep RL studies how trainable systems can learn useful policies from delayed, sparse, and changing feedback instead of fixed labeled datasets.
Reinforcement learning for reasoning
Reasoning-focused RL turns model behavior into an optimization target, raising questions about rewards, emergence, distillation, and where process supervision matters.
Explainable AI as model evaluation
Interpretability matters most when accuracy is not enough: deployment, trust, scientific discovery, fairness, and debugging all require sharper explanations.
Feature attribution and explanation skepticism
LIME, SHAP, saliency maps, and their critiques form a useful testbed for asking whether explanations are faithful, stable, and actually useful to users.
Representation analysis and probing
Representation analysis asks what information is encoded inside a model, how it changes across layers, and whether probes reveal structure or only probe capacity.
Mechanistic interpretability
Mechanistic interpretability is the attempt to reverse-engineer neural networks as computation: circuits, features, induction heads, superposition, and scalable tooling.
Agent evaluation, tool use, and verification
Agent research should be measured by accuracy, cost, latency, repeatability, tool behavior, and whether verification loops catch errors before they compound.
Routing, model fusion, and efficient specialization
Mixture-of-experts, cascades, model merging, task arithmetic, and speculative decoding all ask how to activate only the computation a task deserves.
Contact
Email is the best way to reach me. I usually reply within a week.
- Email praneeth[dot]suresh[dot]s [at] gmail[dot]com
- GitHub github.com/Praneeth-Suresh
- LinkedIn linkedin.com/in/praneeth-suresh
- CV Two-page PDF
What I am open to
- Research conversations about interpretability, model evaluation, and coding-agent reliability.
- Internships and engineering roles in AI tooling, applied ML, and developer tools.
- Feedback on Beryl, SIAP, or a note on this site, including corrections.
- NUS AI Society collaborations: talks, workshops, and partners.
What helps
Say who you are, what you are working on, and what kind of reply would be useful. For a correction, include the page and the claim.