Devanshu Biswas - DEV Community

Devanshu Biswas

Aug 24

Nine Regexes Catch 80% of Injection Attempts and Exactly 0% Once Four Latin Letters Are Swapped for Cyrillic Ones

#ai #llm #security #opensource

3 min read

Devanshu Biswas

Aug 24

Not One Cost Ceiling in the Sweep Improves Cost Per Delivered Task, and Seven Make It Worse: 3,918 Cents Against 1,046

#ai #architecture #python #opensource

3 min read

Devanshu Biswas

Aug 24

A Signed, Id-Bound Approval Token Still Delivers Nine Unauthorised Emails, Because Yes Is Six Bits and It Carries Three

#ai #python #productivity #opensource

3 min read

Devanshu Biswas

Aug 24

Same Corpus, Same Edits, Same Bill: Stale Answers Go 9.1 to 57.3 Per Thousand While the Dashboard Improves

#ai #llm #architecture #database

3 min read

Devanshu Biswas

Aug 24

ECE-15 Agrees With One Bin to 1.9e-16 Before Temperature Scaling, Then Reports a 40% Drop Where the Truth Is 16%

#machinelearning #deeplearning #python #datascience

3 min read

Devanshu Biswas

Aug 24

Squeezing the Scores Moves the Measurable Ceiling From 0.5914 to 0.9480 and the True One Not at All, to 2.2e-16

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 24

One Unchanged W/"abc" Returns 304, 412, 200 and 206, Decided by the Field Name, the Method and an Unrelated Neighbour

#webdev #http #javascript #programming

3 min read

Devanshu Biswas

Aug 24

Deleting One .filter() Buys 30.15 Points for Zero Tokens While the Escape Convention Adds 0.69 Points of Silent Corruption

#ai #llm #prompting #programming

3 min read

Devanshu Biswas

Aug 24

Base-62 Order Keys Cross the Case Boundary on the Second Key They Mint, so a Case-Insensitive ORDER BY Loses 100% of Boards

#javascript #webdev #frontend #database

3 min read

Devanshu Biswas

Aug 24

Nim-Values Wrong About 100% of Positions Win 525 of 525 Games; Ones Wrong About Only 3.09% Win 8 of 719

#javascript #algorithms #gamedev #math

3 min read

Devanshu Biswas

Aug 23

The Only Sentence That Puts Goa Into Arc Rector's Memory Store Is the One Asking For It To Be Forgotten

#ai #llm #python #opensource

3 min read

Devanshu Biswas

Aug 23

A 10-Tick Dip Could Only Fail 600 Requests. Unbudgeted Retries Failed 4,543, and 2,989 After the Cause Had Cleared

#ai #architecture #python #opensource

3 min read

Devanshu Biswas

Aug 23

Truncation Scores 1.00 on Every Metric a Repurposer Can Compute About Itself, and Keeps 0% of the Conclusion

#ai #python #writing #opensource

3 min read

Devanshu Biswas

Aug 23

Identical Model, Prompt and Bill: a Strict Tool Contract Ships 18 Wrong Answers Per Thousand, a Permissive One 240

#ai #llm #architecture #programming

3 min read

Devanshu Biswas

Aug 23

Four of Eight KV-Cache Bugs Are Bit-Identical at Step 1, So the Check Everybody Writes Has Recall Exactly 0.500

#machinelearning #deeplearning #llm #python

3 min read

Devanshu Biswas

Aug 23

The Same 95% Interval Covers the Truth 95.3% of the Time in One World and 0.0% in Another, on Byte-Identical Data

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 23

faß.de and FAẞ.DE Resolve to Two Different Servers, Because Case Folding Runs Before the Deviation Rule

#webdev #security #javascript #programming

3 min read

Devanshu Biswas

Aug 23

Deleting Two Thirds of Your Prompt Hits the Same 183 of 14,400 Cases as the Careful Fix, at 74.5 Tokens Against 229.7

#ai #llm #prompting #machinelearning

3 min read

Devanshu Biswas

Aug 23

All Seven Fill-Handle Implementations Pass on 100.0% of Clean Progressions; Two of Them Are Wrong on 99.3% of Real Columns

#javascript #webdev #frontend #ux

3 min read

Devanshu Biswas

Aug 23

Spreading Evenly Scores Exactly 0 Against a Perfect Blotto Bot and 1 Against One That Merely Watched You

#javascript #algorithms #gamedev #math

3 min read

Devanshu Biswas

Aug 22

Tic-Tac-Toe Is Still a Draw on 20 of 513 Bankrolls; the Other 493 Are Decided Before a Mark Is Placed

#javascript #algorithms #gamedev #probability

3 min read

Devanshu Biswas

Aug 22

A Naive PDF Extractor Welds 18 Sentences Across the Gutter, and Six Whitespace Cuts Remove All Eighteen

#ai #llm #python #opensource

3 min read

Devanshu Biswas

Aug 22

A Payload Hash Plus a Per-Attempt Request ID Is Byte-For-Byte Identical to Having No Idempotency At All

#ai #python #architecture #opensource

3 min read

Devanshu Biswas

Aug 22

Replay Was Meant to Be the Filter and Filters Nothing: 36 of 36 Mined Skills Verify, Including a Chimera

#ai #python #llm #programming

3 min read

Devanshu Biswas

Aug 22

The Same Stream Leaks 5.6 Responses Per Thousand to a Reader and 212 to a Machine, From Reading Speed Alone

#ai #llm #architecture #webdev

3 min read

Devanshu Biswas

Aug 22

The Exposure Gap Ranks Scheduled Sampling's Winners Above Its Losers at AUC 0.529, Against 0.500 for Chance

#machinelearning #deeplearning #python #llm

3 min read

Devanshu Biswas

Aug 22

The Same 95% Interval Covers the Truth 93.6% of the Time and 0.0% of the Time, From Bit-Identical Data

#machinelearning #datascience #python #statistics

3 min read

Devanshu Biswas

Aug 22

The BYxxx-Just-Filters Model Is Exactly Right on All 200 DAILY Rules and Wrong on 276 of 300 YEARLY Ones

#webdev #javascript #programming #datetime

3 min read

Devanshu Biswas

Aug 22

Counting the Lines and Tagging the Answers Both Score 74.0% at k=16, at 92.2 Against 55.4 Tokens per Item

#ai #llm #prompting #machinelearning

3 min read

Devanshu Biswas

Aug 22

Six of Seven Diff Implementations Pass the Only Test Anyone Writes, Including One That Retypes the File at 10.86x

#javascript #webdev #frontend #ux

3 min read

Devanshu Biswas

Aug 21

A Cross-Encoder Reads the Query and the Document Together, Which Your Embedding Model Structurally Cannot

#ai #llm #python #opensource

3 min read

Devanshu Biswas

Aug 21

The Highest-Scoring Duplicate Pair in My Corpus Is Not the Same Story, So No Threshold Reaches Zero

#ai #python #llm #programming

3 min read

Devanshu Biswas

Aug 21

Re-embedding in Traffic Order Beat the Clever Two-Encoder Merge at the Same 30% Budget, 0.803 to 0.673

#ai #llm #machinelearning #architecture

3 min read

Devanshu Biswas

Aug 21

Deleting Half the Gradient of a Tied Embedding Is Exactly Correct on 87.5% of Its Rows

#machinelearning #deeplearning #python #llm

3 min read

Devanshu Biswas

Aug 21

Two Datasets With Bit-Identical Drift Scores Differ 436x in What the Drift Actually Cost

#machinelearning #datascience #python #statistics

3 min read

Devanshu Biswas

Aug 21

Set-Cookie: Secure=1 Is a Cookie Named Secure, and It Travels Over Plain HTTP

#webdev #javascript #security #programming

3 min read

Devanshu Biswas

Aug 21

Taking the Longest Branch Scores 76.2% Against CoT-Decoding's 75.9%, So the Logprobs Bought Nothing

#ai #llm #prompting #machinelearning

3 min read

Devanshu Biswas

Aug 21

Four of the Six Wrong Calendar Layouts Produce Zero Overlapping Boxes, Which Is the Only Thing Anyone Tests

#javascript #webdev #frontend #ux

3 min read

Devanshu Biswas

Aug 21

The Same Graph Under 720 Numberings: 240 Where My Robber Survives Everything, 240 Where It Loses Everything

#javascript #algorithms #gamedev #probability

3 min read

Devanshu Biswas

Aug 19

Your Vector Database Is Three Decisions, and Only One of Them Changes Your Results

#ai #llm #python #opensource

2 min read

Devanshu Biswas

Aug 19

A Pixel Diff Cannot Converge, So Your Screenshot-to-Code Loop Is Following a Flat Signal

#ai #python #llm #programming

3 min read

Devanshu Biswas

Aug 19

A Calibrated 70% Model Beats a Badly Calibrated 80% One in a Cascade, at the Same Cost

#ai #llm #machinelearning #architecture

2 min read

Devanshu Biswas

Aug 19

With Right Padding a Causal Mask Already Excludes Padding. With Left Padding It Excludes Nothing

#machinelearning #deeplearning #python #llm

2 min read

Devanshu Biswas

Aug 19

Selecting Features Before the Split Gave Me 73.7% Accuracy on Data With No Signal in It

#machinelearning #datascience #python #statistics

2 min read

Devanshu Biswas

Aug 19

Adding a Googlebot Block to robots.txt Can Open Your Whole Site to Googlebot

#webdev #seo #javascript #programming

2 min read

Devanshu Biswas

Aug 19

Spotlighting Cut Prompt Injection From 62% to 26%. It Did Not Reach Zero, and It Cannot

#ai #llm #security #prompting

2 min read

Devanshu Biswas

Aug 19

Your Marquee Selection Has Four Bugs and All of Them Keep the Rectangle Inside the Canvas

#javascript #webdev #frontend #ux

2 min read

Devanshu Biswas

Aug 19

Ones Are Wild, So p Is 1/3 and Not 1/6 — and That One Number Decides Whether Your Bot Bluffs

#javascript #algorithms #gamedev #probability

3 min read

Devanshu Biswas

Aug 19

Your Agent Framework Is a Control-Flow Choice, and Three of the Four Options Are Not Frameworks

#ai #llm #python #opensource

4 min read

Devanshu Biswas

Aug 19

My Coding Agent Cannot Fix the Bug, Notices After Two Tries, and Says So. That Is the Feature

#ai #python #llm #programming

4 min read

Devanshu Biswas

Aug 19

I Built Three Models Before I Could Measure Whether a Chain of Thought Was Doing Anything

#ai #llm #machinelearning #programming

3 min read

Devanshu Biswas

Aug 19

Under Adam, Your L2 Penalty Is Divided by the Gradient. The Biggest Gradients Get the Least Decay

#machinelearning #deeplearning #python #programming

3 min read

Devanshu Biswas

Aug 19

The Bootstrap Covers the Mean 94.5% of the Time and the Maximum 0%. I Measured Both

#machinelearning #datascience #statistics #python

3 min read

Devanshu Biswas

Aug 19

strict-dynamic Discards 'self' Too, and My CSP Evaluator Did Not Know That

#webdev #security #javascript #programming

3 min read

Devanshu Biswas

Aug 19

Contextual Calibration Removed 100% of My Prompt-Ordering Effect. That Was the Correct Answer

#ai #llm #machinelearning #prompting

3 min read

Devanshu Biswas

Aug 19

Your Aspect-Locked Image Cropper Has a Bug, and Both of Its Clamps Are Individually Correct

#javascript #webdev #frontend #ux

3 min read

Devanshu Biswas

Aug 19

Minimax Does Not Apply to Backgammon. The Node Between You and Your Opponent Is a Die

#javascript #algorithms #gamedev #programming

3 min read

Devanshu Biswas

Aug 17

Llama Is Not Open Source. The Clause That Decides It Has a Number In It: 700 Million Users

#ai #llm #opensource #machinelearning

5 min read

Devanshu Biswas

Aug 17

My Multi-Agent Crew Found 2.11x as Many Facts as One Agent. It Also Cost 4.5x and Was Less Precise

#ai #python #llm #programming

5 min read

Devanshu Biswas

Aug 17

I Tried to Reproduce Attention Sinks and Got Nothing. The Null Result Is the Interesting Part

#ai #llm #machinelearning #programming

3 min read

Writing Debut

Want to connect with Devanshu Biswas?