Devanshu Biswas - DEV Community

Devanshu Biswas

Aug 4

The human-in-the-loop agent: an 8B model proposes, Python pauses the risky calls, a human decides

#ai #llm #agents #python

3 min read

Devanshu Biswas

Aug 4

One Helm template, eight Deployments: how range over values.yaml collapses your k8s YAML

#kubernetes #devops #springboot #java

3 min read

Devanshu Biswas

Aug 4

Lost in the middle: why a 100k-token context window is not 100k tokens of attention

#ai #llm #machinelearning #rag

3 min read

Devanshu Biswas

Aug 4

ALiBi: give a Transformer position sense with one line, and let it test longer than it trained

#machinelearning #deeplearning #transformers #python

3 min read

Devanshu Biswas

Aug 4

Random projections: compress 10,000 dimensions to 30 by multiplying with random noise

#machinelearning #datascience #python #algorithms

3 min read

Devanshu Biswas

Aug 4

ULID from scratch: 128 bits that sort by time, and why UUIDv4 never can

#javascript #webdev #programming #node

3 min read

Devanshu Biswas

Aug 4

Verbalized confidence: make an LLM state how sure it is, then check if 90% really means 90%

#ai #llm #promptengineering #machinelearning

3 min read

Devanshu Biswas

Aug 4

Build iOS Settings from scratch: a data-driven list, a real toggle, and a push-navigation stack

#css #javascript #webdev #html

3 min read

Devanshu Biswas

Aug 4

Yahtzee in 180 lines: how one counts[1..6] array scores all 13 categories

#javascript #gamedev #beginners #webdev

3 min read

Devanshu Biswas

Aug 3

A memory agent that remembers you after the process dies — short-term buffer + vector recall + compression, no vector DB

#ai #llm #agents #python

2 min read

Devanshu Biswas

Aug 3

From docker-compose to Kubernetes: 8 Spring Boot services, probes that gate rollout, and a test that reads the manifests

#kubernetes #springboot #java #devops

2 min read

Devanshu Biswas

Aug 3

Test-time compute: the two-line math behind majority vote, best-of-N, and "thinking longer"

#ai #llm #machinelearning #python

3 min read

Devanshu Biswas

Aug 3

Gated Linear Units explained: the multiplicative valve behind LLaMA and PaLM's SwiGLU FFN

#machinelearning #deeplearning #transformers #python

3 min read

Devanshu Biswas

Aug 3

The Theil-Sen estimator: fit a robust regression line with nothing but a median

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 3

Writing an HTML entity encoder from scratch: which chars really need escaping, and why you iterate by code point

#javascript #webdev #html #programming

3 min read

Devanshu Biswas

Aug 3

Chain-of-Draft: keep the reasoning, drop the narration, and cut ~80% of your reasoning tokens

#ai #llm #promptengineering #chatgpt

3 min read

Devanshu Biswas

Aug 3

The ellipsis truncation algorithm: how a pagination bar stays a fixed width no matter the page count

#css #javascript #webdev #html

3 min read

Devanshu Biswas

Aug 3

Building Bomberman from a grid of integers: a 2-second fuse, a cross-blast raycast, and BFS hunters

#javascript #gamedev #algorithms #beginners

3 min read

Devanshu Biswas

Aug 2

Why components passed as children don't re-render with the parent (a free re-render boundary in React)

#react #typescript #webdev #performance

3 min read

Devanshu Biswas

Aug 2

I built a binary heap visualizer: bubble-up, sift-down, and O(n) heapify with the array and tree side by side

#algorithms #javascript #computerscience #datastructures

3 min read

Devanshu Biswas

Aug 2

A multi-tool orchestrator on llama-3.1-8b — capability routing, deny-by-default permissions, and a 2.97 parallel speed-up

#ai #llm #agents #python

3 min read

Devanshu Biswas

Aug 2

Eight services and six pieces of infra — one docker compose up brings the whole stack up in dependency order

#docker #springboot #java #devops

3 min read

Devanshu Biswas

Aug 2

GraphRAG answers connect-the-dots questions flat vector RAG can't — by walking a knowledge graph, not a pile of chunks

#ai #llm #rag #machinelearning

3 min read

Devanshu Biswas

Aug 2

CLIP learns from captions, not labels — two encoders, one shared space, and an N N matrix whose diagonal must win

#machinelearning #deeplearning #computervision #ai

3 min read

Devanshu Biswas

Aug 2

Huber regression bends but doesn't break — a quadratic core, linear tails, and one line of IRLS reweighting

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 2

A JSON Schema validator is one recursive function — schema, value, path — that calls itself and collects every violation

#javascript #webdev #json #tools

3 min read

Devanshu Biswas

Aug 2

A chat only grows but the context window is fixed — context management is choosing what survives

#ai #llm #promptengineering #chatgpt

1

3 min read

Devanshu Biswas

Aug 2

A chip input is a text box and a list fused into one control — built from one plain input and an array of strings

#css #javascript #webdev #html

1

3 min read

Devanshu Biswas

Aug 2

Mastermind needs no rules engine — just two arrays of four numbers and one duplicate-safe two-pass scoring trick

#javascript #gamedev #algorithms #beginners

3 min read

Devanshu Biswas

Aug 2

I watched an 8B model solve a multi-step task, then degrade honestly on a 503 — a ReAct agent with real guardrails

#ai #llm #agents #python

3 min read

Devanshu Biswas

Aug 2

A whole notification feature added by only listening — event-driven decoupling means new consumers are free

#springboot #java #kafka #backend

3 min read

Devanshu Biswas

Aug 2

M N integrations collapse to M+N — the Model Context Protocol is USB-C for AI tools, and it's just JSON-RPC underneath

#ai #llm #mcp #agents

3 min read

Devanshu Biswas

Aug 2

A Vision Transformer makes an image look like a sentence — 16 16 patches become tokens and every patch attends to every other

#machinelearning #deeplearning #computervision #python

3 min read

Devanshu Biswas

Aug 2

Least squares has no defence against a single bad point — RANSAC just lets the outliers fail to vote

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 2

A fluid font size is just a line through two points — clamp() and a little algebra retire the whole media-query stack

#css #webdev #frontend #javascript

3 min read

Devanshu Biswas

Aug 2

The system prompt is a separate channel — role, rules, format and guardrails the model reads before the first user turn

#ai #llm #promptengineering #chatgpt

3 min read

Devanshu Biswas

Aug 2

A FAB is not a special widget — it is one plain button pinned to a corner, and one boolean opens the whole speed-dial

#css #javascript #webdev #html

3 min read

Devanshu Biswas

Aug 2

There is no level in the code — only five plain arrays and a clock, and a whole tower-defense game falls out of them

#javascript #gamedev #canvas #beginners

3 min read

Devanshu Biswas

Jul 31

Render props vs hooks: the same logic three ways, and the wrapper hell that made hooks win

#react #typescript #webdev #javascript

2 min read

Devanshu Biswas

Jul 31

I built a union-find visualizer: watch union by size and path compression flatten a tree in real time

#algorithms #javascript #computerscience #datastructures

2 min read

Devanshu Biswas

Jul 31

A RAG agent that refuses to bluff — inline [n] citations when grounded, a low-confidence flag when not, a fallback when blank

#ai #llm #rag #python

3 min read

Devanshu Biswas

Jul 31

A URL rule guards the door by verb and path — it can't say "only your own order", so authorization moves onto the method

#springboot #java #security #backend

3 min read

Devanshu Biswas

Jul 31

Don't ask the model for valid JSON — delete the wrong tokens, force every illegal logit to , and it's valid by construction

#ai #llm #json #machinelearning

3 min read

Devanshu Biswas

Jul 31

fp16 doesn't just lose precision, it loses range — gradients underflow to zero, and loss scaling borrows exponent headroom

#machinelearning #deeplearning #python #gpu

3 min read

Devanshu Biswas

Jul 31

A histogram is the crudest density estimate there is — drop a smooth bump on every point, sum them, and bandwidth is everything

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Jul 31

A URL is a tiny structured record, not a string — parse the frame with the platform, hand-roll only the query, rebuild it

#javascript #webdev #tools #programming

3 min read

Devanshu Biswas

Jul 31

Auto-CoT writes its own chain-of-thought demos — cluster the questions, pick one per cluster, and diversity beats similarity

#ai #llm #promptengineering #machinelearning

3 min read

Devanshu Biswas

Jul 31

Confetti isn't an animation — it's an array of paper scraps, each a bag of numbers one loop nudges under gravity every frame

#javascript #css #webdev #animation

3 min read

Devanshu Biswas

Jul 31

There is no flock in the code — only an array of points, and three little vectors per bird make a whole murmuration appear

#javascript #gamedev #canvas #simulation

3 min read

Devanshu Biswas

Jul 31

OrderHub Day 39: distributed tracing + Loki — one trace-id across four services, into every log line, 135 tests

#springboot #java #observability #tracing

3 min read

Devanshu Biswas

Jul 31

Semantic caching keys on meaning, not the exact string — embed the query, cosine-match the cache, HIT above a threshold

#ai #llm #machinelearning #performance

3 min read

Devanshu Biswas

Jul 31

Gradient accumulation fakes a big batch on small memory — .grad already sums, so scale each micro-batch by 1/N and step once

#machinelearning #deeplearning #python #ai

3 min read

Devanshu Biswas

Jul 31

Spectral clustering: cluster by the graph Laplacian's eigenvectors, not by distance — non-convex shapes unfold into blobs

#machinelearning #datascience #python #ai

3 min read

Devanshu Biswas

Jul 31

A JSONPath engine is two moves — parse the path into a list of selectors, then pipe a node-set of {path, value} through each

#javascript #webdev #json #programming

3 min read

Devanshu Biswas

Jul 31

Buffer of Thoughts: keep a library of reusable thought-templates, not past answers — retrieve, instantiate, distil back

#ai #llm #promptengineering #beginners

3 min read

Devanshu Biswas

Jul 31

A bottom sheet is one translateY — snap points are fractions of the frame, and release projects the flick to the nearest one

#css #webdev #frontend #javascript

3 min read

Devanshu Biswas

Jul 31

A centipede is just an array ordered head-first — every segment copies the cell ahead, and a bolt splices the array in two

#javascript #gamedev #canvas #beginners

3 min read

Devanshu Biswas

Jul 29

OrderHub Day 38: a Grafana dashboard provisioned from git — rate() and histogram_quantile() over the Day-37 meters, 131 tests

#springboot #java #observability #grafana

1

3 min read

Devanshu Biswas

Jul 29

Sliding-window attention: a banded mask turns O(N ) into O(N·W), and stacking L layers buys an effective field of ~L·W

#ai #llm #machinelearning #transformers

1

3 min read

Devanshu Biswas

Jul 29

Gradient checkpointing: keep N activations, recompute the rest — memory falls from O(N) to O( N) for one extra forward pass

#machinelearning #deeplearning #ai #python

3 min read

Writing Debut

Want to connect with Devanshu Biswas?