About
I'm a software engineer who is passionate about building out large-scale web services and…
Activity
-
🚀 Docling just crossed a major milestone: 500,000 package downloads per day! 🚀 What started as an open-source effort to make document…
🚀 Docling just crossed a major milestone: 500,000 package downloads per day! 🚀 What started as an open-source effort to make document…
Liked by Oleg Silkin
-
Chips are the fuel for AI. By using AI to design, optimize, and automate chip design, we can close the recursive self-improving loop between AI and…
Chips are the fuel for AI. By using AI to design, optimize, and automate chip design, we can close the recursive self-improving loop between AI and…
Liked by Oleg Silkin
-
Coding agents are great at building things, but surprisingly bad at figuring out what to build. Hao Wang's latest work introduces SpecBench, a…
Coding agents are great at building things, but surprisingly bad at figuring out what to build. Hao Wang's latest work introduces SpecBench, a…
Shared by Oleg Silkin
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
Framingham State University
-
Activities and Societies: Elected President at Math Club
Emphasis in Tensor Calculus and its application in Artificial Intelligence
-
-
Licenses & Certifications
Projects
-
HappyMail
See projectSmall program which will look through emails you recieve and move the negative, non-constructive emails to a seperate inbox so users don't experience the anxiety of rejection when applying for job positions
-
DogeAI
- Present
See projectImage Classifier that detects pictures of doge interfaced with the Twitter API optimized for the most positive user interactions with NLP automated server management technology
-
CppNN
- Present
See projectExtensible, High-Level C++ Neural Network API designed for maximum user-friendliness and little requirement for user's to need to know what's being done under-the-hood.
-
LBRY Comments
-
See projectMicroservice within the LBRY application stack which manages comments.
The service provides the functionality for users to create, edit, or delete their own comments, as well as spam-filtering and moderation.
This is achieved by maintaining a database which tracks users, comments, and the content on which the comments are made. -
Grafton Atlantic - Transaction Generator
-
Created a software product for a client which tracks user sessions in a bank application. When a user starts a session within the application, a transaction is created and its ID is linked to the user's account number.
We developed the program within an agile workcycle, using a scrum board to assign and manage our tasks. Making use of continuous integration tools, we quickly iterated through designs and created a clear picture of what our client was looking for.
Other creators -
-
ReFill
-
See projectModernized profanity filter concept; Unwanted phrases are carefully replaced with ones that are generated to sound nonsensical.
For instance: "wow, what an absolute schmuck" would get replaced with "wow, what an absolute gummy bear." -
Machine Learning with 2048
-
See projectSimple Neural Genetic Algorithm that learns to play a version of 2048 using NVIDIA's CUDA platform.
Honors & Awards
-
Dean's List, Fall 2020
Margaret Carroll, Ph.D. Dean of the College of Science, Technology, Engineering, and Mathematics (STEM)
Completed a full program with at least three (3) credit-bearing undergraduate courses in the
Day and/or Continuing Education Divisions, earned a semester Grade Point Average (GPA) of at least
3.30, and had no incomplete grades. -
President's List, Fall 2020
F. Javier Cevallos, PhD, President
Carried a program of at least three 3 course-credits while earning a semester Grade Point Average (GPA) of 3.75 or higher.
-
Dean's List, Spring 2020
Margaret Carroll, Ph.D. Dean of the College of Science, Technology, Engineering, and Mathematics (STEM)
Completed a full program with at least three (3) credit-bearing undergraduate courses in the
Day and/or Continuing Education Divisions, earned a semester Grade Point Average (GPA) of at least
3.30, and had no incomplete grades. -
President's List, Spring 2020
F. Javier Cevallos, PhD
The President’s List recognizes those students who have carried a program of at least three (3) course credits while earning a semester cumulative Grade Point Average (GPA) of 3.30 or higher for
three consecutive semesters. -
Dean's List, Fall 2019
Margaret Carroll, Ph.D. Dean of the College of Science, Technology, Engineering, and Mathematics (STEM)
Completed a full program with at least three (3) credit-bearing undergraduate courses in the
Day and/or Continuing Education Divisions, earned a semester Grade Point Average (GPA) of at least
3.30, and had no incomplete grades. -
Dean's List, Spring 2019
Margaret Carroll, Ph.D. Dean of the College of Science, Technology, Engineering, and Mathematics (STEM)
Completed a full program with at least three (3) credit-bearing undergraduate courses in the
Day and/or Continuing Education Divisions, earned a semester Grade Point Average (GPA) of at least
3.30, and had no incomplete grades. -
President's List, Spring 2018
-
-
AP Scholar with Honor Award
AP CollegeBoard
Recognizes exemplary college-level achievement, for attaining an
average score of at least 3.25 on all AP Exams taken, and scores of 3 or higher
on four or more exams.
Languages
-
Russian
Native or bilingual proficiency
-
Chinese
Limited working proficiency
More activity by Oleg
-
Congratulations to Charles H. Bennett and IBM on the Turing award - widely called Nobel Prize of computer science! An amazing milestone on the…
Congratulations to Charles H. Bennett and IBM on the Turing award - widely called Nobel Prize of computer science! An amazing milestone on the…
Liked by Oleg Silkin
-
Join our Product Management team live on February 24th for our first roadmap session of 2026. We’ll unpack the new capabilities in Red Hat AI and…
Join our Product Management team live on February 24th for our first roadmap session of 2026. We’ll unpack the new capabilities in Red Hat AI and…
Liked by Oleg Silkin
-
Qwen just released 4 new Qwen3.5 Small models!🔥 Run 0.8B, 2B and 4B on your phone via Unsloth AI GGUFs. Run 9B on 6GB RAM. The vision reasoning…
Qwen just released 4 new Qwen3.5 Small models!🔥 Run 0.8B, 2B and 4B on your phone via Unsloth AI GGUFs. Run 9B on 6GB RAM. The vision reasoning…
Liked by Oleg Silkin
Other similar profiles
Explore more posts
-
Hameer Singh
About 2 months ago, I was in a KT meeting, confidently nodding through a Git discussion. Ten minutes later — guess who was secretly Googling “rebase vs merge”? 👀 Truth is, even experienced devs don’t have every syntax or flag memorized. We remember where to look, not what to recall. Same with Regex or Docker commands — it’s less about knowing everything, and more about knowing how to find, apply, and adapt fast. So next time you hit Google mid-task, don’t feel guilty — it’s not a weakness. It’s a developer’s superpower. 💪 #DeveloperLife #TechHumor #SoftwareEngineering #GrowthMindset #DevCommunity
-
Pavan Pothnis
I've always found myself switching to Claude models when working through coding problems. There's just something about how Claude handles context and works its way through large code bases and complex logic. With Claude Code, that workflow has become ridiculously smooth. Also been experimenting with OpenAI's Codex (GPT-5 powered) and the results are way better than previous versions. The code generation quality has taken a real jump. Now here's where it gets interesting. You keep hearing stories about developers opening multiple coding agents, feeding them the same task, and manually cross-referencing outputs to find the best solution. Turns out there's an easier way. By integrating Codex as a skill inside Claude Code, you get insights from both agents in a single terminal session when you need. No tab switching, no copy-pasting, no manual comparison. Claude Code orchestrates the workflow while Codex brings its strengths. Best of both worlds in one window. Claude OpenAI
2 Comments -
Dev Parikh
The hard part of fine-tuning is rarely the model. I put together a Phi-3 QLoRA pipeline that goes from Dolly 15k training data to a FastAPI inference service, with MLflow tracking in the middle and optional ONNX export for deployment. What looked simple on paper got tricky once memory and serving constraints showed up. 4-bit training with bitsandbytes wants a real CUDA setup, and the "just export to ONNX" step is fragile enough that fallback paths matter. Build shape: - TRL SFTTrainer + PEFT LoRA on microsoft/Phi-3-mini-4k-instruct - Databricks Dolly 15k preprocessing into chat-format messages - MLflow run tracking plus optional registry promotion for merged weights - FastAPI endpoints for /health, /generate, and base-vs-finetuned /compare - Optimum ONNX Runtime export for a lighter inference backend One decision I now feel strongly about: treat training and serving as one system early. It is easy to get a decent loss curve and still ship a model artifact that is painful to load, compare, or operate. The most useful fix was boring but practical: save both adapter and merged checkpoints, then design the API to run PyTorch and ONNX backends behind the same contract. That gave me a safer path when ONNX export behaved badly, and a cleaner path when it worked. Current config is intentionally small for smoke tests (max_steps 50), so the next step is pushing longer runs and stronger evals beyond eval_loss. If you have tuned small LLMs recently, what broke first for you: training memory, export reliability, or inference plumbing? #LLM #MLOps #FineTuning #FastAPI #ONNX
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content