Anjana - AI & Tech
Открыть в Telegram
🏁 Dive into the world of AI with Anju! 🔍 Al Insights & Updates: This channel simplifies Al, shares practical resources, and keeps you updated on: ✅ Al, ML, and automation explained ✅ Global Al trends and innovations ✅ Educational content for learners
БольшеСтрана не указанаКатегория не указана
6 135
Подписчики
+16924 часа
+9737 дней
+2 18830 дней
Загружаем данные
Похожие каналы
Нет данных
Возникли проблемы? Пожалуйста, обновите страницу или обратитесь к нашему support-менеджеру .
Облако тегов
Нет данных
Возникли проблемы? Пожалуйста, обновите страницу или обратитесь к нашему support-менеджеру .
Входящие и исходящие упоминания
---
---
---
---
---
---
Привлечение подписчиков
октябрь '26окт. '26
октябрь '26
+6 139
в 1 каналах
| Дата | Привлечение подписчиков | Упоминания | Каналы | |
| 26 октября | +126 | |||
| 25 октября | +169 | |||
| 24 октября | +284 | |||
| 23 октября | +283 | |||
| 22 октября | +63 | |||
| 21 октября | +73 | |||
| 20 октября | +57 | |||
| 19 октября | +48 | |||
| 18 октября | +62 | |||
| 17 октября | +60 | |||
| 16 октября | +79 | |||
| 15 октября | +101 | |||
| 14 октября | +99 | |||
| 13 октября | +140 | |||
| 12 октября | +454 | |||
| 11 октября | +108 | |||
| 10 октября | +112 |
Посты канала
| 2 | Asked in a Swiggy AI/ML Engineer interview 👇
"It's raining heavily in Bangalore. Swiggy's AI predicted the order to surge 45 minutes before it happened and repositioned delivery partners in advance. Design that system."
Here's how you crack it step by step:
Step 1: By the time orders spike, it is too late.
Delivery partners need to already be positioned when demand hits. Predict rain's effect before it changes behavior, not react once orders climb.
Step 2: Use weather nowcasting as the earliest signal.
Pull precipitation probability and radar showing storms approaching, not current weather. Nowcasting predicts rain 30 to 60 minutes out, the earliest leading indicator, well before order patterns shift.
Step 3: Apply a historical rain-to-demand correlation model.
A light drizzle at 3pm behaves differently than heavy rain at 7pm dinner time. Train a model on past rain events and order volume, by time of day and zone.
Step 4: Catch early behavioral signals as rain starts.
App open rate rising and browsing without immediate orders are intent signals available minutes before any order is placed. These sit between the forecast and the actual surge.
Step 5: Forecast demand at the zone level, not city-wide.
Combine weather forecast, historical correlation and behavioral signals into a per-zone prediction. Bangalore's zones have different rain sensitivity based on density. City-wide forecasting misses this.
Step 6: Calculate the supply-demand gap and reposition.
Compare predicted surge per zone against current partner availability. Where demand outstrips supply, send repositioning incentives to partners in adjacent zones, with a pre-announced bonus for moving before the surge hits.
Step 7: Re-forecast continuously and close the loop.
The prediction updates every few minutes. After the event, compare the actual surge to predicted and refine the model. Track whether repositioning reduced delivery times. | 468 |
| 3 |  Нет текста... | 461 |
| 4 |  🚀 5 Unsupervised Machine Learning Algorithms You Should Know! 👇
https://www.instagram.com/reel/De52nqGSTSg/?rpxt=anpmenFob25kNjIz | 946 |
| 5 | Unsupervised Machine Learning Algorithms Complete Guide 👆 | 1 263 |
| 6 | 5 Most Important Unsupervised Machine Learning Algorithms You Should Know 👆 | 1 266 |
| 7 |  Нет текста... | 1 467 |
| 8 | Asked in an Uber AI/ML Engineer interview 👇
Uber’s AI predicts that ride demand in a particular area will increase significantly in the next 20 minutes. How would you design an AI system that predicts demand and proactively moves drivers to that area?
Prediction is the easy half. The value is in the allocation, and most candidates stop before they get there.
1. Forecast both sides, per cell per bucket
- H3 hexagons × 5-minute buckets, forecasting rides requested and drivers available
- Features: the same hour last week, weather, events, traffic, and live app-open sessions
- Short horizons need current state; long horizons need seasonality
2. What matters is the gap
- Rank cells by expected unmet demand, which is demand minus reachable supply
- A surging cell that already has forty idle drivers needs nothing from you
3. Turn the forecast into assignments
- For each driver, the expected value of moving versus staying, minus travel time and fuel
- Solve it as min-cost flow across drivers and cells, not as a greedy nudge per driver
- Greedy sends everyone to the same hex and creates the opposite problem
4. Drivers are people, not units of supply
- Incentives and heatmaps, never orders — model the probability they accept
- Respect end-of-shift, home direction and rest; an ignored nudge is worse than none
5. Do not stampede
- Move a fraction of the nearby drivers, then re-solve a few minutes later
- Closed-loop correction beats one confident dispatch
6. Evaluate on the marketplace
- Geo-based A/B tests, because moving drivers changes the outcome for everyone in the city
- Track unfulfilled requests, pickup ETA and driver earnings per hour together
The principle: A demand forecast is only worth what the allocation on top of it earns. Optimise the assignment, not the RMSE.
💬 Incentives or dispatch — which moves drivers better in your experience? | 1 516 |
| 9 |  Нет текста... | 1 156 |
| 10 | 10 Essential Agentic AI concepts to know in 2026.pdf | 2 212 |
| 11 |  Нет текста... | 2 303 |
| 12 |  Нет текста... | 2 606 |
| 13 | Asked in a Goldman Sachs AI Engineer interview 👇
Your chatbot receives both simple FAQ questions and complex financial-analysis requests. Would you send both requests to the same LLM? If not, how would you design the routing architecture?
No — say that in the first sentence. Sending "what are your business hours" to a frontier reasoning model pays 30x the price and 10x the latency for an answer a cached string already had.
1. Route on task, not on message length
- Cheap tier: FAQ, greetings, status lookups — a small model, or a cache hit
- Mid tier: retrieval-grounded answers over policy and product documents
- Heavy tier: multi-step analysis with tools, real numbers and citations
2. Build the classifier as a cascade
- Exact and semantic cache first, so repeated questions never reach a model at all
- Then rules and patterns for the obvious cases
- Then a small classifier, ~50ms, returning intent plus a confidence score
- Only genuinely ambiguous traffic pays for an LLM router
3. Let the cheap path escalate
- Low confidence, a refusal, or a failed validation promotes the request one tier up
- Escalating the few is far cheaper than routing everything to the top
4. Give every tier its own contract
- Separate prompts, tools, timeouts, token budgets and eval sets
- Analysis gets a calculator and a market-data tool; FAQ gets neither
5. Guard the expensive tier
- Per-user rate limits and a hard spend ceiling
- Queue long analyses and stream progress instead of holding a request open
- Numbers come from tools, never from the model's head
6. Measure the router itself
- Misroute rate in both directions, cost per resolved conversation, p95 per tier
- A router that never escalates is quietly degrading quality
The principle: one model for every request is either too expensive for the easy questions or too weak for the hard ones. Routing is where cost and quality are actually decided.
💬 Rules, a classifier, or an LLM router — what is in production for you? | 2 251 |
| 14 |  Нет текста... | 1 886 |
| 15 | Complete Machine Learning Engineer Roadmap - Download👆 | 1 967 |
| 16 |  🚀 Follow this roadmap to become a Machine Learning Engineer in 2026👇
https://www.instagram.com/reel/Des_6UOSplh/?exln=MWJvaXoxY2wxY3VpdA== | 2 164 |
| 17 | Complete Machine Learning Engineer Roadmap and Resources👇️️ | 1 902 |
| 18 | Asked in a CRED AI Engineer interview 👇
A payment platform processes millions of transactions every minute. How would you design an AI system that detects potentially fraudulent transactions in real time while minimizing false positives?
Do the arithmetic on the false positives. At a million transactions a minute, a 0.1% false-positive rate blocks a thousand genuine payments a minute. Accuracy is not the objective here — cost is.
1. Fix the latency budget first
- The decision sits inline with authorisation: 50-100ms
- Features precomputed in a feature store and read by key, never joined at request time
- Streaming aggregates (Flink, Kafka) keep the counters fresh
2. Two models, not one
- A gradient-boosted tree on tabular features scores every transaction in single-digit ms
- The expensive graph or sequence model runs only on the risky slice
3. Features carry the signal
- Velocity: count and amount over 1 minute, 1 hour and 24 hours per card and device
- Deviation from this user's own baseline, not the population average
- Device fingerprint, IP, and geo-velocity — impossible travel is a gift
4. Score the network, not the transaction
- Shared devices, accounts and payout targets expose rings a per-transaction model never sees
- Graph features (component size, distance to a known bad node) move precision the most
5. Minimise false positives by acting in bands
- One threshold is a blunt instrument: approve, step up to OTP or 3DS, then block
- A step-up turns a would-be false positive into a little friction
- Tune for precision at a fixed recall, and price both errors in rupees
6. Design for labels that arrive late
- Chargebacks land 30-90 days later, so retraining always trails the fraudster
- Keep a small unblocked holdout to measure what you missed
- Shadow-score challengers on live traffic before promoting them | 1 793 |
| 19 |  Нет текста... | 1 722 |
| 20 |  Нет текста... | 2 019 |
